Use cases
Explore how enterprises apply AI and automation across financial services, technology, mobility, media, and retail.
23 use cases
Fintech and Banking
Stripe developed an AI-driven fraud prevention system called Radar that assesses over 1,000 characteristics of a transaction in under 100 milliseconds to accurately detect fraud while minimizing false positives. Radar employs techniques like deep neural networks, feature engineering, and leveraging the breadth of the Stripe network.
Fintech and Banking
Monzo developed an AI-driven system to optimize marketing messages for users, using uplift models to predict the additional value a user would gain from a campaign compared to not receiving one. The system employs multi-objective optimization to determine the most optimal campaign targeting for each user.
Fintech and Banking
Monzo used topic modelling to analyze customer Pot names and categorize them into themes to gain insights into how customers are using Pots and what they are saving for. The analysis revealed that 30% of Pots are for generic saving, 15% are for travel, and seasonal trends like increased 'life events' Pots near the end of the year for Christmas. Emojis were also leveraged to further categorize Pots and gain insights into specific destinations customers are saving for travel.
Fintech and Banking
Nubank developed Precog, an AI system that encodes customer events (app clicks, transactions) as sequences of symbols to predict customer support needs in real-time. It uses techniques like embeddings and sequence models to understand and predict customer needs.
Fintech and Banking
PayPal developed a two-layer ensemble classifier framework to make progressive predictions on the likelihood of opportunities in the sales pipeline being closed or lost. The first layer is a binary classifier that generates an initial propensity score when an opportunity first appears. The second layer incorporates dynamic features like duration in the pipeline to update the propensity score over time.
Fintech and Banking
Didact AI developed an ML-powered stock picking engine that generated outstanding returns, steadily beating the S&P 500 for over a year on a weekly basis. The engine made progressive predictions on the likelihood of opportunities in the sales pipeline being closed or lost using a two-layer ensemble classifier framework.
Fintech and Banking
Brex developed an automated system called 'Auto Merchant Cat' to identify merchant names, websites, and categories from credit card transaction data. The system uses Google Places, machine learning, and Amazon Mechanical Turk to categorize merchants into 48 Brex categories for business accounting.
Fintech and Banking
Capital One developed a machine learning-based suspicious activity monitoring system to identify and prioritize money laundering activity. The system uses a random forest model to analyze customer and transaction attributes and generate a score indicating the likelihood of suspicious activity.
Fintech and Banking
Capital One developed a real-time failure detection and remediation platform that uses machine learning to monitor application data, detect anomalies, and identify the root cause of failures. The system integrates monitoring data, uses specialized anomaly detection models, and leverages dependency graphs and causal models to quickly identify the root cause and notify engineering teams, reducing incident resolution time by up to 50%.
Fintech and Banking
PayPal developed a deep learning-based framework to optimize actions for domain-specific objectives. The framework uses custom loss functions to address specific business needs. For example, in product propensity modeling, a custom loss function is used to maximize sales while considering the cost of reaching out to customers.
Fintech and Banking
Capital One developed a machine learning-based system within the Eno browser extension to accurately detect payment pages and locate payment fields, making the virtual card number (VCN) experience more seamless and secure for customers.
Fintech and Banking
PayPal developed machine learning models to predict issuer declines and optimize payment authorization rates. The models use features like past decline ratios, issuer patterns, and transaction timing to accurately forecast if a payment will be declined.
Fintech and Banking
Stripe developed a machine learning-based system to detect and prevent fraudulent transactions. The system uses supervised learning models to classify transactions as fraudulent or legitimate based on features like transaction history, user behavior, and device information.
Fintech and Banking
PayPal developed a real-time graph database platform to detect and prevent fraudulent transactions by analyzing relationships between entities like buyers, sellers, and transactions. The platform uses graph analysis techniques like page rank, connected components, and centrality algorithms to identify fraud rings, compromised accounts, and risky buyer-seller interactions.
Fintech and Banking
Nubank developed causal inference techniques to estimate the 'treatment effect' or sensitivity of each customer to actions like credit limit increases, interest rate changes, or marketing emails. This allows optimizing decisions to maximize business metrics like long-term customer value, conversion rates, and profitability.
Fintech and Banking
Adyen developed a contextual multi-armed bandit system to optimize payment conversion rates by selecting the best payment optimizations for each transaction. The system uses an XGBoost classifier as an oracle to predict authorization probabilities and an epsilon-greedy policy for action selection.
Fintech and Banking
PayPal developed an AI-driven system using multi-task deep learning to detect various fraud patterns simultaneously while optimizing fraud catch rates across different domains. The system balances effective fraud detection with minimizing false positives, employing techniques like hard parameter sharing, multiple shared layers, clustering constraints, cross-stitch units, and robust feature learning using stacked denoising autoencoders.
Fintech and Banking
Identifying and grouping similar transactions to detect fraud rings effectively.
Fintech and Banking
The use case involves utilizing Large Language Models (LLMs) to automatically tag sensitive data within the delivery and mobility, and banking and finance industries. This approach aims to classify data entities at scale, ensuring that sensitive information is accurately identified and tagged.
Fintech and Banking
The use case involves utilizing unsupervised graph anomaly detection techniques to identify new types of fraudulent behaviors in the delivery and mobility, and banking and finance industries.
Fintech and Banking
The use case aims to enhance marketing efforts by creating scalable lookalike audiences for advertisers, allowing for more effective targeting of potential customers.
Fintech and Banking
The use case involves scaling marketing efforts for merchants through targeted and intelligent promotional campaigns to optimize their impact.
Fintech and Banking
The use case focuses on detecting fraud using graph-based models, which can effectively identify complex relationships and patterns indicative of fraudulent activities.
45 use cases
The description typically follows the title and provides an overview of what the AI system is designed to do. It should outline its capabilities, purpose, and possibly some key features or benefits.
The description typically follows the title and provides an overview of what the AI system (GitHub Copilot in this case) is designed to do. It should explain its functionality, purpose, and how it integrates within GitHub's ecosystem.
The description typically follows the title and provides an overview of the AI system's role or purpose within the context of the blog post. It should explain what the AI system is designed to do and possibly its significance or impact.
Typically follows the title and provides an overview of the AI system or concept discussed in the article. It should explain what the system does, its purpose in the context of recommender systems, and possibly its benefits or applications.
Typically found after the title, the description provides an overview of the AI system or research topic. It should explain what salient object segmentation is, its importance, and its potential applications.
Typically found near the beginning of the blog post, the description provides an overview of the feature or AI system. It should explain what query suggestions are, how they enhance user experience, and their benefits in search functionality.
The abstract of the paper, usually found near the top of the arXiv page, provides a summary of the research. It should briefly describe the problem studied, the methodology used, and the key findings or contributions.
Typically found near the beginning of the blog post, the description provides an overview of the AI system or feature. It should explain what Einstein Search Answers is, how it functions, and its role within Salesforce's AI offerings.
Typically found near the beginning of the blog post, the description provides an overview of the topic. It should explain what GitHub Copilot is used for in enterprise applications, its benefits, and possibly its challenges or considerations.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain what BOBA is (likely an acronym or a specific system), its purpose, and its role in software development or another relevant field.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain the problem addressed (identifying date formats in file names), the approach taken using machine learning, and the potential benefits or applications.
Typically found near the beginning of the blog post, the description provides an overview of the topic. It should explain what the AI Summarist does, how it enhances productivity within Slack, and its potential benefits for users.
Typically found near the beginning of the blog post, the description provides an overview of the topic. It should explain the importance of diversity and explainability in enterprise app recommendation systems, the challenges addressed, and the objectives of the research or discussion.
Typically found near the beginning of the blog post, the description provides an overview of the topic. It should explain how generative AI is used to enhance compliance in software development, the challenges addressed, and the benefits of using AI in this context.
Typically found near the beginning of the blog post, the description provides an overview of the topic. It should explain how auto-generated summaries work in Google Docs, the technology behind it, and the benefits it brings to users.
Typically found near the beginning of the blog post, the description provides an overview of the topic. It should explain how machine learning is used to detect security vulnerabilities, the methodology employed, and the significance of using AI in this context.
Typically found near the beginning of the blog post, the description provides an overview of the feature. It should explain how conversation summaries work in Google Chat, their purpose, and how they benefit users.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain the approach used to estimate past promotions' profitability, the challenges encountered, and the insights gained from the analysis.
Typically found near the beginning of the blog post, the description provides an overview of the topic. It should explain the technology and methodology behind how Grammarly generates suggestions, including the linguistic and machine learning aspects.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain Peloton's journey in integrating machine learning into their operations, the specific applications of ML discussed, and the benefits derived.
Typically found near the beginning of the blog post, the description provides an overview of the topic. It should explain the specific areas or processes within engineering that Grammarly focuses on innovating, the challenges addressed, and the innovations implemented.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain the approach or methodology used to optimize test feedback in machine learning, the challenges addressed, and the outcomes achieved.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain the approach or techniques used to detect and prevent invite spam on Slack using machine learning, including the challenges and benefits.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain how Dropbox implements image search functionality, the algorithms used, and the benefits to users.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain the specific applications of machine learning in customer support, the challenges addressed, and the benefits achieved.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain how Google Photos utilizes AI for reminiscing features, the technology behind it, and how users benefit from these capabilities.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain how topic modeling is applied at Microsoft for customer support, the methodology used, and the insights gained from the analysis.
Typically found near the beginning of the page, the description provides an overview of the research topic. It should explain the methods and technologies Apple uses for recognizing people in photos, including the challenges and advancements in the field.
Typically found near the beginning of the blog post, the description provides an overview of the topic. It should explain the approach or methodology used by Datto to predict hard drive failures using machine learning, the types of data analyzed, and the outcomes achieved.
Typically found near the beginning of the blog post, the description provides an overview of the topic. It should explain how Dropbox leverages machine learning for optimizing payments, the specific challenges addressed, and the improvements achieved.
Typically found near the beginning of the blog post, the description provides an overview of the topic. It should explain the strategies, technologies, or methodologies Grammarly employs to enhance reader engagement and attention.
Typically found near the beginning of the research page, the description provides an overview of the topic. It should explain the approach, methodology, and objectives of Apple's research in interpretable adaptive optimization.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain the methods, techniques, or algorithms used by Slack for email classification, as well as the purpose and benefits of the classification system.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain the purpose and methodology behind Grammarly's decision to use tagging for corrections instead of rewriting text.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain the problem statement, the approach using graph-based techniques, and the significance of fast layout detection.
Typically found near the beginning of the blog post, the description provides an overview of the technology or feature discussed. It should explain how Google implements grammar correction as users type on Pixel devices, the underlying technology or algorithms, and the user experience improvements.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain the concept of adversarial training in grammatical error correction, the challenges it addresses, and the benefits of this approach.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain the approach Microsoft took in leveraging ML and heuristic labeling, the specific challenges addressed in cloud migration, and the outcomes achieved.
Typically found near the beginning of the article, the description provides an overview of the feature. It should explain the purpose of "Good First Issues," how GitHub identified and categorized these issues, and the benefits for open-source contributors.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain the specific applications of machine learning in Firefox testing, the challenges addressed, and the improvements achieved.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain how Duolingo integrates AI across different aspects of its app, such as language learning algorithms, personalized learning experiences, and user engagement strategies.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain the approach Mozilla takes in using machine learning for bug triaging, the challenges addressed, and the benefits of this approach.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain the purpose of the content suggestion system, how Dropbox employs machine learning algorithms for this purpose, and the benefits for users.
Typically found near the beginning of the article, the description provides an overview of the topic. It should explain how ZoomInfo measures and interprets machine learning accuracy scores, the importance of accuracy in their applications, and potentially any specific methodologies or tools used.
Typically found near the beginning of the research page, the description provides an overview of the topic. It should explain the challenges of identifying languages from short pieces of text, the methods or algorithms used by Apple, and the applications or benefits of accurate language identification.
47 use cases
Swiggy's AI system aims to predict and understand customer preferences to enhance their food ordering experience. The system leverages advanced machine learning techniques to analyze customer behavior, preferences, and historical data to recommend the most relevant and desirable food options to users.
Uber aims to optimize the supply-demand balance at airports by forecasting demand and estimated time to request (ETR) for the FIFO driver queue. This helps drivers decide whether to wait at the airport or reposition themselves based on real-time data on demand and supply.
Foodpanda's menu ranking system aims to optimize the presentation of menu items to maximize conversion rates and enhance the user experience. The system leverages data science and machine learning to rank menu items based on various factors such as popularity, user preferences, and historical data.
DoorDash developed a machine learning (ML) system to minimize the wait times for Dashers (delivery drivers) by predicting the optimal time to release food orders to the kitchen. The goal was to synchronize food preparation with Dasher arrival times, thereby reducing idle times and enhancing overall efficiency.
Lyft's recommendation system leverages machine learning to optimize various aspects of the user experience, such as suggesting destinations based on historical data, optimizing routes, and coordinating rider pickups efficiently. The system adapts to real-time data, improving ETA accuracy and routing based on current conditions and user behavior patterns.
The AI system addresses the challenge of accurately tracking merchants' operational status on the DoorDash platform. Inaccurate reporting of store closures can lead to canceled orders, customer dissatisfaction, and lost revenue for merchants. The system employs a machine learning model to predict whether a store is closed based on Dashers' reported closures, reducing the need for manual verification and improving overall efficiency.
The article describes DoorDash's development of an ensemble learning model for time series forecasting to predict demand accurately. This system integrates multiple machine learning models to improve forecasting accuracy, handle non-linear relationships, and adjust to varying data patterns. The goal is to enhance delivery logistics by predicting order volumes more precisely.
DoorDash developed a cascade machine learning (ML) approach to enhance their holiday demand predictions. The system addresses the challenge of significant fluctuations in order volumes during holidays by using a series of ML models, each refining the predictions of the previous one. This layered approach ensures more accurate and reliable forecasts, helping DoorDash optimize resource allocation and maintain service quality during peak times.
BlaBlaCar utilizes machine learning techniques to detect and prevent fraudulent activities on their platform. The system aims to solve specific problems such as identifying suspicious user behaviors and preventing unauthorized transactions, enhancing platform security using a data-driven approach.
Risk Entity Watch utilizes unsupervised machine learning to scrutinize transaction patterns and user behaviors, identifying deviations that potentially indicate fraud. This proactive approach enhances the security and integrity of Uber's services.
Swiggy's AI system utilizes generative technologies to enhance operations, predict customer preferences, and manage supply chain logistics more efficiently, improving overall food preparation processes.
The AI system at Delivery Hero is designed to enhance user experience through personalization. It utilizes both implicit and explicit feedback from users to tailor recommendations, ensuring each customer receives the most relevant suggestions based on their previous interactions and preferences.
This section would detail how Swiggy's AI system uses historical data, real-time traffic conditions, restaurant preparation times, and other logistics factors to accurately predict the delivery time right when an order is placed.
This section would describe the architecture and workings of the machine learning pipeline designed to detect and prevent fraud. It probably includes details on how data is processed, features are engineered, and machine learning models are applied to identify fraudulent behavior.
Delivery Hero uses machine learning to personalize restaurant rankings for new users. By addressing the cold start problem, they enhance user experience by matching new customers with restaurants based on factors like cuisine, budget, and location.
This AI system enhances user experience by personalizing restaurant recommendations using both explicit and implicit customer data, analyzing preferences based on previous interactions.
The system uses embeddings to improve search relevance by capturing the semantic relationships between products and search terms, allowing for more accurate and contextually appropriate search results on Instacart's platform.
This system analyzes restaurant data and user feedback to categorize cuisines effectively, handling the subjectivity in how different people describe food. It likely uses natural language processing to interpret diverse culinary descriptions.
The system uses data science techniques to identify patterns and predictors of customer churn. It likely utilizes historical data, purchase patterns, and customer feedback to forecast retention risks and suggest proactive measures.
This system would use statistical models and machine learning to predict order volume at Grubhub, helping manage resources and optimize delivery times. It likely analyzes historical data, seasonal trends, and other variables to make accurate forecasts.
RADAR integrates advanced machine learning algorithms with expert human analysis to detect and respond to fraud dynamically. It analyzes fine-grained market segments, identifying potential fraud early and enabling rapid response to minimize losses.
Tensoba enhances ETA predictions for GoFood by analyzing historical data and various real-time factors like traffic, order details, and restaurant distance. This system aims to improve customer satisfaction by providing more accurate delivery time estimates.
The description would explain how the AI system addresses the cold start problem in recipe recommendations, likely by using techniques to gather initial user preferences or leveraging similar user data.
CCG uses machine learning and linear programming to optimize push notification schedules for Uber users, ensuring messages are timely and relevant. This system improves user engagement by carefully selecting which notifications to send based on various operational constraints.
BAIR addresses the challenge of personalizing recommendations by considering both user feedback and the uncertainty in early-stage user preferences. It uses a multi-armed bandit approach to dynamically adjust recommendations based on user interactions.
UR combines deep learning for exploitation (leveraging known preferences) and exploration (discovering new options), enhancing user engagement by balancing relevance and novelty.
MxModel leverages machine learning to identify potential merchants that will add value to the DoorDash platform, focusing on those that match consumer demand and preferences. It analyzes merchant characteristics and market data to predict success on the platform.
An enhanced search system designed to broaden the scope of DoorDash’s product search functionality beyond traditional delivery options. This system incorporates advanced algorithms to facilitate a more extensive and user-friendly search experience. It adapts to user preferences and regional trends, ensuring relevant and diverse product listings.
The article likely describes how Tensoba uses real-time data and machine learning models to accurately predict the time it takes for food to move from restaurants to customers. This system may incorporate factors like traffic, order complexity, and restaurant efficiency.
Lyft's pricing system utilizes deep learning, reinforcement learning, and linear programming to dynamically adjust fares. It aims to balance supply and demand, considering factors like traffic, events, and typical ride patterns to optimize pricing strategies for both drivers and passengers.
This system would utilize deep learning techniques to analyze and compare the text of an address with its geographic coordinates. It aims to identify mismatches that could indicate errors in data entry or fraudulent activities.
The system enhances safety and security on the Uber platform by verifying the identities of riders using real-time document checks. It uses advanced document processing technologies and human reviews to ensure the accuracy and authenticity of IDs presented by users.
The article would describe how the AI system uses machine learning to predict service times based on various factors such as customer behavior, historical data, and perhaps environmental conditions, improving operational efficiency.
The article would describe how the AI system uses causal inference techniques to forecast and analyze the effect of specific actions or events on Lyft’s service metrics. This could involve understanding how changes in pricing, promotions, or external events like weather affect demand.
This part likely delves deeper into the methodologies used in causal forecasting at Lyft, possibly discussing specific case studies or examples where causal inference has significantly impacted business decisions.
The description would cover how the system uses historical order data, seasonal trends, and possibly real-time events to predict future order volumes, helping Grubhub optimize its logistics and marketing strategies.
The tool applies machine learning to enhance the auditing process, specifically targeting issues like vendor verification and transaction validation. It aims to identify discrepancies and potential frauds that traditional methods might miss.
The Rouxcommender system uses machine learning to tailor recipe recommendations to individual customer preferences. It likely analyzes user data and feedback to continuously improve its accuracy and relevance.
This system likely uses machine learning techniques to rank restaurants based on various criteria such as customer preferences, order history, and ratings, aiming to enhance user experience by presenting the most relevant choices first.
The system uses transfer learning with pre-trained computer vision models, allowing for rapid deployment across different image recognition tasks. It is optimized for speed and efficiency, enabling real-time data processing and quick adaptation to various business needs.
This would be a system designed specifically for improving the ranking of dishes in search results, possibly named something indicative of its function, such as 'DishRank.' The AI system uses deep learning techniques to rank dishes more effectively based on various factors like popularity, relevance to search terms, and user preferences.
DefraudNet utilizes a weak supervision approach, leveraging limited labeled data to train models that can detect complex patterns indicative of fraudulent activities within the online food delivery sector.
DeepRed handles the complex dispatch problem by balancing efficient route planning and timely deliveries. It employs ML models to predict various aspects of delivery and optimization algorithms to make strategic dispatch decisions, enhancing Dasher efficiency and customer satisfaction.
This would likely be a system designed for predicting the travel distance for two-wheelers, essential for logistics and delivery optimization at Swiggy. The system uses machine learning models to predict two-wheeler travel distances accurately, taking into account various factors like road conditions, traffic, and historical data.
The AI system uses machine learning to predict service times accurately, processing various kinds of operational data to enhance efficiency and customer experience.
CartoBERT leverages deep learning to analyze booking texts and identify meaningful pickup point names. It improves upon traditional statistical models by understanding contextual nuances, thereby enhancing the accuracy and relevance of the pickup point suggestions.
The system discusses how Lyft uses attention mechanisms within machine learning to enhance the accuracy of predicting a user's intended destination, improving route planning and customer experience.
52 use cases
The system utilizes machine learning to analyze user data and tailor marketing strategies, enhancing customer engagement by delivering personalized content and recommendations.
The system uses machine learning and natural language processing to enable users to search for specific content within videos, improving the viewer's ability to find and access relevant segments.
The system uses machine learning algorithms to analyze user viewing patterns and preferences to create personalized video clips (sizzles) that aim to increase engagement and viewing time.
The article likely describes how the system leverages neural network algorithms to analyze user preferences and listening history to recommend music tracks that users are likely to enjoy.
This section would likely describe how the AI system uses machine learning techniques to analyze audio signals and accurately identify segments containing speech and music, enhancing content categorization and accessibility features.
The system likely uses vector space modeling to analyze user preferences combined with opinion mining to understand sentiments expressed in user feedback, aiming to deliver more personalized content recommendations.
The description might cover how the system uses user interactions, preferences, and historical data to recommend recipes that align with the user's tastes and cooking habits.
A machine learning system designed to create short, engaging previews for podcasts to enhance content discovery on Spotify.
This section likely describes how Vimeo implemented generative AI to improve their customer support services. The system uses advanced natural language processing models to understand customer queries and provide accurate, helpful responses, thus elevating the overall user experience.
The system uses NLP to semantically match user queries with podcast episodes, rather than relying solely on exact keyword matches, then ranks the results based on relevance and user experience.
The system uses sophisticated algorithms to analyze user interactions and preferences, making real-time recommendations through the Alexa voice assistant. By gathering contextual information during conversations, it delivers music suggestions tailored to user moods and activities.
The system provides accurate user forecasts by processing large datasets through multiple pipelines, including mature markets with robust historical data and new markets using proxy-based models. It enables both on-demand and weekly automated forecasts, incorporating advanced hyperparameter tuning and visualization tools to support business decisions.
The system uses sophisticated algorithms to analyze user interactions and preferences, making real-time recommendations through the Alexa voice assistant. By gathering contextual information during conversations, it delivers music suggestions tailored to user moods and activities.
This system leverages machine learning to analyze user behavior patterns and detect anomalies that may indicate fraudulent activities, such as account sharing or unauthorized access. By continuously learning from data, the system can adapt to new fraud tactics and provide real-time alerts.
The system uses machine learning to analyze user behavior and determine the optimal time to present the paywall, aiming to convert free readers into subscribers without negatively impacting user experience.
This system applies reinforcement learning to balance content recommendations against budgetary limits. It aims to maximize user engagement while adhering to financial constraints, ensuring the most cost-effective allocation of resources.
The system employs advanced neural networks to enhance video quality by reducing compression artifacts and improving visual clarity. This AI-driven approach allows for real-time adjustments to streaming quality, providing a superior viewing experience.
This system utilizes machine learning algorithms to enhance the relevance and diversity of video recommendations. By balancing user preferences with diverse content, the system aims to keep users engaged and satisfied with varied video suggestions.
Gaming
Zynga developed a deep reinforcement learning system to optimize the timing of push notifications for their game Words With Friends, aiming to increase user engagement. The system learns to select the best hour of the day to send notifications to each user based on their historical interactions.
Gaming
King developed a deep learning system called DeepTester to automate playtesting for mobile games. The system uses imitation learning to mimic human player behavior, allowing for large-scale testing of game levels and features.
Social platforms
Twitter's recommendation system distills approximately 500 million daily tweets into a personalized 'For You' timeline. It uses interconnected services and models to extract latent information from tweet, user, and engagement data. The system aims to deliver relevant tweets by predicting user interactions and trending content.
Social platforms
LinkedIn's system improves the relevance of the homepage feed by using large corpus sparse ID embeddings. This approach captures rich contextual information about user interactions and content, leading to better personalization and user engagement.
Social platforms
This AI system focuses on enhancing user engagement on online platforms by utilizing generative AI combined with rejection sampling. The goal is to generate content that not only appears informative or well-written but also resonates with users to increase activities such as clicks and interactions. The system is particularly employed in the context of generating email notification subject lines, which showed significant improvements in engagement metrics, including a 1% increase in session and a 0.4% increase in weekly active users.
Social platforms
The AI system focuses on improving the relevance and accuracy of Pinterest's closeup recommendation ranker. This system uses machine learning models to rank the content shown to users when they interact closely with items, such as zooming in on a pin. The improvements aim to enhance user experience by providing more personalized and contextually relevant recommendations.
Social platforms
Code Llama is a state-of-the-art large language model developed by Meta, designed specifically for coding tasks. It can generate and understand code from text prompts, supporting multiple programming languages such as Python, C++, Java, PHP, Typescript, C#, and Bash. The model aims to assist developers in writing, completing, and debugging code more efficiently, enhancing both productivity and learning.
Social platforms
SeamlessM4T is a foundational multilingual and multitask model for speech and text translation and transcription. It supports automatic speech recognition for nearly 100 languages, speech-to-text and text-to-speech translation for nearly 100 languages, and speech-to-speech translation for 35+ languages. The model leverages large datasets and advanced architecture to provide seamless translation capabilities, improving communication across different languages.
Social platforms
Pinterest has developed a recommender system for its Homefeed, leveraging a model known as the Pinnability model. This system uses pretrained user embeddings and sequences of short-term user interests to provide relevant recommendations. The model incorporates real-time user action signals to boost engagement by adapting to the immediate preferences and activities of users.
Social platforms
The Constructive Conversations Reminder is an AI-powered feature by Nextdoor that aims to foster respectful and productive conversations. It anticipates when a discussion may become heated and prompts users with tips on maintaining a positive tone. This feature helps mitigate negative responses and promotes civility by considering the overall context of conversations.
Social platforms
eBay developed a Multi-Relevance Ranking Model to enhance the recommendation of similar items on its platform. The system leverages machine learning to incorporate various purchase intent signals beyond simple purchase data, such as clicks, adding items to the cart, and other user actions, to improve the relevance of recommended items.
Social platforms
Yelp transitioned from a collaborative filtering approach using matrix factorization to a hybrid recommendation model. The new model integrates content-based features (e.g., business reviews, ratings) with interaction data to enhance recommendations, overcoming limitations for users with few interactions.
Social platforms
Yelp transitioned from a collaborative filtering approach using matrix factorization to a hybrid recommendation model. The new model integrates content-based features (e.g., business reviews, ratings) with interaction data to enhance recommendations. This hybrid model overcomes limitations of matrix factorization, particularly for users with few interactions ('tail' users), providing more accurate and personalized recommendations.
Social platforms
Meta uses machine learning and causal inference to personalize and manage Instagram notifications, particularly daily digest push notifications. By predicting user engagement and optimizing notification delivery, they aim to enhance user experience by reducing unnecessary notifications, especially for highly active users who are likely to see the content organically.
Social platforms
LinkedIn revamped its Post search system to enhance relevance and user experience. The new system decouples from the Feed ecosystem and integrates multiple machine learning models to balance query relevance, personalization, post engagement, and recency.
Social platforms
The system serves ads to users by filtering eligible ads, generating candidate ads, using heavy ML models to score them, running an auction to select the highest value ad, and using a feedback loop for continuous improvement.
Social platforms
LinkedIn developed Project Account Prioritizer, an ML-based system, to predict upsell and churn for its SaaS products. This system helps sales teams prioritize accounts. CrystalCandle is an explainable AI tool that provides narrative insights to explain the model's predictions.
Social platforms
LinkedIn transitioned from a GLMix model to a deep-learning-based system for predicting ads click-through-rate (CTR). This new model architecture includes three towers—deep, wide, and shallow—that together improve prediction accuracy and model relevance.
Social platforms
Twitter developed a two-tower model for account recommendations, enhancing personalization by using embeddings to reflect consumer and producer behaviors. This model helps improve the relevance of suggested accounts by incorporating various features into a unified framework.
Social platforms
Ocelot is LinkedIn's platform for scaling observational causal inference. It helps estimate the causal impact of product changes on key metrics when A/B testing is infeasible or too costly. The platform leverages methods like Coarsened Exact Matching, Doubly Robust estimators, Instrumental variables, Fixed effects models, and Bayesian Structured Time Series to handle observational data.
Social platforms
This system utilizes learned embeddings to provide personalized recommendations for users. By leveraging a sophisticated embedding model, the system can understand and predict user preferences and behaviors, enabling more relevant and accurate recommendations within the Fishbowl platform. The embeddings are designed to capture intricate relationships between different user attributes and interaction patterns, facilitating a deeper understanding of user needs.
Social platforms
This AI system leverages learned embeddings to personalize recommendations within Fishbowl, a professional networking community. The system focuses on using text embeddings to understand and rank the relevance of posts, thereby enhancing user experience by providing more pertinent content recommendations based on the user's interaction and preferences.
Social platforms
LinkedIn introduced a new approach to enhancing job matching by leveraging machine-learned activity features. This system constructs 'activity embeddings' for users based on their job-seeking activities, allowing for more personalized and relevant job recommendations. It moves beyond traditional content-based recommendations by incorporating recent job-seeking behaviors, thus adapting to users' current job preferences and activities.
Social platforms
This system utilizes multitask learning to improve AI model performance across various LinkedIn services, including job recommendations and job search. By training models on related tasks simultaneously, it allows for shared learnings and better overall results, addressing challenges such as label distribution differences and feature discrepancies among tasks.
Social platforms
This system leverages machine learning to predict the CPU time and peak memory usage of SQL queries. It was designed to enhance query scheduling, provide resource usage estimates to users, and support elastic scaling. The system utilizes historical SQL query logs to train models that forecast resource consumption without depending on traditional SQL engines or query plans.
Social platforms
This AI system is designed to optimize ad delivery on Pinterest by leveraging machine learning algorithms. The goal is to ensure that users are presented with ads that are highly relevant to their interests, thus improving user engagement and advertising effectiveness. The system uses various data points and user behaviors to build a comprehensive understanding of user preferences.
Social platforms
This system utilizes clustering and automated rule creation to fight spam on Pinterest. By grouping similar content and identifying patterns, the system can detect and filter out spam more efficiently. This method allows Pinterest to maintain the integrity and quality of content on its platform, providing a better user experience.
Social platforms
This system is designed to moderate message content across multiple languages at scale. It leverages machine learning models to detect inappropriate or harmful content in various languages, ensuring that the platform remains safe and welcoming for all users. The system can handle the complexities of different languages and dialects, providing accurate moderation regardless of the language used.
Social platforms
This system utilizes machine learning to optimize the delivery of notifications, ensuring that users receive relevant and timely information. The AI models analyze user behavior and engagement patterns to tailor notifications, enhancing user experience and keeping neighbors informed about important updates and events in their local community.
Social platforms
Pinterest's Comment Moderation System leverages machine learning to maintain a healthy comment ecosystem on the platform. The system detects and removes harmful comments, ensuring that interactions remain positive and constructive. By analyzing patterns in user behavior and comment content, the AI can identify and flag inappropriate comments effectively.
Social platforms
Bumble's Multilingual Message Content Moderation System uses advanced machine learning techniques to moderate messages across multiple languages at scale. It ensures that the platform remains safe and welcoming by detecting and managing harmful content. The system is designed to handle the unique challenges posed by different languages and cultural contexts.
Social platforms
This system leverages advanced machine learning techniques, such as the Two Towers neural networks, to provide personalized content recommendations on the Instagram Explore page. The system's design aims to scale and adapt to the massive volume of content and user interactions, ensuring relevant and engaging content is surfaced to users.
Social platforms
The description usually follows the introduction of the AI system. It explains what the AI system does, its purpose, functionality, and possibly its impact or significance within Pinterest's ecosystem. This system likely recommends advertisers or ad content to users based on their behavior and preferences.
Social platforms
The description usually follows the introduction of the AI system. It should explain what the AI system does, its purpose (in this case, related to news feed ranking), and possibly its impact or significance within Facebook's platform. This system likely ranks news feed content based on user engagement and preferences.
70 use cases
E-commerce and Retail
Walmart developed a machine learning system to generate personalized outfit recommendations, helping customers discover complete, stylized outfits built around an item they are considering. The system uses a combination of algorithmic look generation and look coverage expansion to provide style-aware, personalized recommendations across Walmart's large product catalog.
E-commerce and Retail
Stitch Fix developed a generative AI system that combines algorithmic text generation with human expert oversight to create engaging ad headlines and detailed product descriptions. The system uses GPT-3 for ad headlines and a fine-tuned language model for product descriptions, with human experts reviewing and refining the output.
E-commerce and Retail
Zillow developed a deep learning system called the Neural Zestimate to provide automated home value estimates for over 100 million homes in the U.S. The system uses a single national-scale model that incorporates property data, sales transactions, tax assessments, and location information to generate accurate and responsive home value estimates.
E-commerce and Retail
Wayfair developed a machine learning system to predict delivery dates for products, improving accuracy and customer experience. The system uses gradient boosted decision trees (GBTs) with quantile regression to estimate delivery times end-to-end, considering factors like supplier performance, distance, and temporal features.
E-commerce and Retail
Zalando developed a deep learning-based forecasting system to predict demand for fashion products. The model addresses challenges specific to online fashion, including high data volume, catalog turnover, and the fixed inventory assumption. It uses a global approach to handle the large and irregular dataset, incorporating price-demand relationships.
E-commerce and Retail
Etsy developed a multi-task canonical ranker to provide personalized product recommendations to their customers. The system uses a single deep learning model to jointly optimize for multiple recommendation tasks, such as related items, complementary items, and personalized ranking. This allows the model to learn shared representations and generalize better.
E-commerce and Retail
Hamlet is a machine learning system developed by Wayfair to differentiate between B2C (consumer) and B2B (business) shoppers on their website. It aims to accurately predict if a customer's order was placed on behalf of a business in order to engage them with Wayfair Professional, the company's B2B offering.
E-commerce and Retail
Melange is a customer journey embedding system developed by Wayfair to improve fraud and policy abuse detection on their platform. It uses machine learning models to analyze user behavior and transactions across multiple dimensions to identify suspicious activity.
E-commerce and Retail
Instacart developed a new real-time item availability prediction system to improve the accuracy and efficiency of their inventory management. The system uses a hierarchical model structure with three components: General, Trending, and Real-Time.
E-commerce and Retail
SpectroBrain is a machine learning system developed by Zillow Rentals to detect spam in voice message recordings and identify offending callers. It uses semi-supervised learning to overcome the challenge of limited labeled training data.
E-commerce and Retail
Instacart developed a contextual bandit (CB) model to personalize search ranking by recommending the best action to take in each observed context. The system uses a hierarchical model structure with three components: General, Trending, and Real-Time.
E-commerce and Retail
Etsy developed adSformers, a system that uses a combination of pretrained representations and learned representations to personalize ads for each user. It captures short-term user behavior sequences and leverages diverse representations including visual, multimodal (AIR), and skip-gram.
E-commerce and Retail
Stitch Fix uses a multi-tiered recommender system to personalize styling recommendations at scale. The scoring module is based on the Client Time Series Model (CTSM), an award-winning sequence-based model that uses temporally masked encoders. As the model size and training data grew, training time increased significantly, making it harder to iterate quickly.
E-commerce and Retail
Griffin is a reinforcement learning system developed by Wayfair to make real-time decisions on which marketing emails and push notifications to send to customers. It aims to optimize customer engagement and reduce unsubscriptions by learning from millions of daily interactions.
E-commerce and Retail
Whatnot developed a centralized rule engine as the foundation for their trust and safety platform, which applies data-informed heuristics to determine policy violations by analyzing data from multiple sources. They use LLMs to broaden the understanding of user interactions and conversations for more nuanced content moderation and violation detection, enabling proactive scam detection by analyzing entire messaging exchanges to detect patterns like asking for card details, pretending to be in urgent need of money, and indicating payment failures. Using automated detection with LLMs, Whatnot is able to proactively detect over 95% of scam attempts within minutes, with 96% precision.
E-commerce and Retail
The rise of COVID-19 led to a significant surge in demand for Instacart's services, as more customers turned to the platform for their grocery shopping needs. This created a logistical challenge for Instacart to determine the physical availability of millions of items across thousands of retail locations, each with their own inventory management systems. The pandemic also caused disruptions in supply chains and labor shortages, resulting in rapidly fluctuating store inventory levels. This made it more difficult for Instacart's machine learning models to accurately predict product availability, leading to a larger discrepancy between predicted availability and actual in-store stock.
E-commerce and Retail
Building an infrastructure to serve Real-time Availability (RTA) predictions required addressing challenges like supporting fetching of scores via RPC, low latency use cases, and high-consistency use cases. Instacart implemented two methods for ingesting scores into DB storage: Full Sync and Lazy Score Refresh. They developed a multi-model experimentation framework with DB column per model, model-column mapping, and experiment-column mapping to accelerate ML model testing. The Deltas Framework applies fixed deltas to base thresholds to compute final thresholds at runtime, offering modularity and enabling easy experimentation of ML models.
E-commerce and Retail
OLX used machine learning to improve delivery time estimation to encourage buyers to use Pay&Ship. The project involved defining the problem and success criteria, exploratory data analysis, baseline solution, A/B testing, training and modeling, productionizing the solution, and identifying next steps. Key findings included the importance of carrier, day of the week, region, and distance from seller to buyer. The Catboost model improved performance by 13.3% compared to the baseline.
E-commerce and Retail
Ava is an internal AI assistant powered by OpenAI’s GPT-4 and GPT-3.5 models. It was initially built as a hackathon project and has since seen accelerated adoption, with over half of Instacart employees using it every month and more than 900 using it every week. Ava’s features include a web interface, conversation search, automatic model upgrades, and a “Prompt Exchange” that allows users to share and reuse prompts. Users use Ava to write, review and debug code, improve comms, learn faster, and build AI-enabled internal tools on top of Instacart’s APIs.
E-commerce and Retail
eBay developed a deep learning-based retrieval system with two main components: 1) Embedding Generation - a neural network generates embeddings for items and users, capturing semantics, visual appearance, and preferences. 2) Retrieval - an ANN index is built using item embeddings. The user embedding is used to query the index and retrieve relevant items. Key innovations include multi-task learning, hierarchical taxonomy, leveraging user interactions, and optimizing the ANN index. The system significantly outperforms traditional methods in relevance and engagement. Future plans include incorporating more data signals, improving embeddings, and optimizing the retrieval infrastructure.
E-commerce and Retail
Walmart developed a unified multi-task model to support virtual assistants for grocery shopping, customer care, and store associates. The model leverages commonalities in retail entities like products and brands across domains. It improves person and time recognition by annotating store live logs using state-of-the-art NER models and labeling functions. The unified model achieves competitive performance on retail entities compared to individual models, while significantly improving person and time recognition. It also shows better generalization on intent classification by leveraging data from multiple domains. Key benefits include reduced maintenance cost, ability to share common entities, and leveraging data and domain knowledge across domains to improve performance.
E-commerce and Retail
Walmart developed an Entity Transformer (ET) neural network architecture to recognize up to 10 product name entities in a single voice ordering utterance. The ET model uses an encoder-decoder architecture with transformer layers to encode the input utterance and a CRF decoder to produce the sequence of entity tag predictions. It leverages dense features from pre-trained language models like conveRT, BERT, and RoBERTa, along with sparse features like one-hot token encodings and character n-gram features. Evaluation on a 1M product name entity dataset showed the best ET model configuration had a 12% performance improvement over non-neural models and outperformed BERT with ET. This helps customers efficiently finalize their shopping cart via voice dialog, improving the overall shopping experience.
E-commerce and Retail
Instacart uses autocomplete to help users express their search intent, explore the product catalog, and build larger baskets. The autocomplete suggestions come from previous customer searches, filtered to remove bad words, non-shopping terms, and plural duplicates. Instacart uses an autocomplete engagement model to rank the suggestions, considering features like conversion rate, skip rate, fuzzy matching, and thumbnails. They also handle misspellings via fuzzy matching and semantic deduplication using search embeddings. A multi-objective ranking model optimizes for both autocomplete engagement and add-to-carts. Leveraging the Instacart catalog helps with the cold start problem for new retailers. Autocomplete has driven a 2.7% increase in add-to-carts per search and 1.6% increase in ad revenue.
E-commerce and Retail
Home Insights are key phrases extracted from property listing descriptions using machine learning and NLP. The pipeline includes keyword extraction using fine-tuned BERT for named entity recognition, filtering improper phrases, and ranking insights by trending searches in each region. Home Insights are displayed on listing detail pages and search results to help users explore and compare homes. Future plans include personalizing insights based on user preferences and enabling search by Home Insights. The pipeline extracts insights covering interior, exterior, renovation, location, and other attributes to provide a rich summary of each home's unique features.
E-commerce and Retail
Auto Trader built a real-time CDP that ingests user behavior tracking events from Snowplow. Events contain identity information like unique user ID and Auto Trader account ID, along with details of the action taken (e.g. viewing a car advert). The CDP is a purpose-built real-time database that stores a unified profile for each customer based on their recent activity. Auto Trader assigns customers to segments based on their preferences, such as fuel type (electric, petrol, diesel) and body type (SUV, hatchback, convertible). They use a simple model based on the proportion of vehicles a customer has already viewed that match each segment to decide when to add them to a segment. A customer can be in at most one fuel type segment. When a customer lands on Auto Trader, their unique user ID is used to fetch their profile from the CDP. The segments are concatenated into the request header and passed to the Vehicle Listing Service (VLS) that provides search results. VLS can then manipulate the search query to prioritize vehicles matching the customer's segments, but only if they haven't already applied a filter for that attribute. Auto Trader runs AB tests to validate the impact of personalizing the Featured listing position. Customers are randomly assigned to a test or control bucket. Test customers in a segment are shown a vehicle matching their preference, while control customers see a non-personalized Featured listing. The test is analyzed only for customers who have a preference segment. AB tests have shown an average uplift of 4.8-7.2% in advert views per session for body type segments, and 2.1-4.1% for fuel type segments.
E-commerce and Retail
Faire uses a feature store to consolidate and serve features for their real-time ranking models. The feature store includes both offline features computed daily in their data warehouse, and online real-time features computed on-the-fly. Offline features are stored in a key-value store for low-latency retrieval, while online features are computed by querying production databases and caching recent user activity. Features are recorded for each ranking impression, enabling straightforward model training using the 'log-and-wait' approach. The feature store provides a centralized registry of all features with metadata, monitoring for missing/drifting features, and a simple API for creating new features. This enables plug-and-play models that can be deployed with just a few lines of code.
E-commerce and Retail
Zillow developed a likely-buyer propensity model to predict how likely a user is to transact on a home in the near future. The model uses past user behavior on Zillow's website and apps as input features, including onsite activity (home viewing, saving, time spent), offsite activity (email/push interactions), and changes in user preferences over time. The target labels are extracted from completed real-estate transactions matched to registered Zillow users. The model is trained as a binary classifier using gradient boosting decision trees. High-intent buyers are identified by users in the top deciles of the model's predicted propensity scores. Key challenges include heavily imbalanced positive/negative labels and only partially known ground truth. The model enables personalizing experiences for high-intent buyers across marketing and product features. Future plans include near-real-time intent classification to serve users based on their evolving preferences.
E-commerce and Retail
Walmart collects sensor data like pressure, fan speed, temperature, defrost status, etc. for refrigeration systems at a 10-minute interval. The data is aggregated to hourly level and imputed for missing values. Forecasting is done using Facebook's Prophet model with outside air temperature as an exogenous feature. Anomalies are detected when forecasted temperature is outside the favorable temperature band for the system type. Anomaly priority is based on magnitude and duration of being outside the band. Prophet's change-point detection is used to identify potential causes of anomalies by looking for changes in feature variables within a 2-3 day window of a temperature pattern change. The framework is deployed on Walmart's cloud platform with daily batch processing of around half a million refrigeration systems. Validation showed 90%+ forecast accuracy for 85% of systems and 67% recall in capturing actual malfunction issues. Future plans include minimizing resolution time, identifying root causes at the mechanical level, and recommending actions to address recurring issues.
E-commerce and Retail
CTSM is Stitch Fix's novel sequence-based recommendation model that uses a Temporally-Masked Encoder (TME) to capture the evolution of client preferences over time. It learns a unified client embedding that is updated with each new interaction, and uses separate mappings to different prediction domains. CTSM reflects client updates instantly across platforms, without requiring complex feature engineering. It models two types of events - Updates (changes to client state) and Targets (outcomes to predict). The TME computes weights only once for the whole sequence to encode updates in a time-safe manner. CTSM has led to improvements in key performance metrics, reduced computing resources, lower maintenance, and accelerated development compared to Stitch Fix's previous recommendation models.
E-commerce and Retail
Zillow's Valuation team aims to enhance the Zestimate valuation model by integrating listing description information. They explored using listing description embeddings as features and leveraging Zillow's Home Insights to extract and categorize phrases, identifying features like swimming pools, decks, and yards not found in structured data. This involves mapping Home Insights tags to a real estate taxonomy, computing embeddings, and using a KNN model to detect features based on similarity thresholds, improving Zestimate accuracy.
E-commerce and Retail
Walmart developed a scalable product recommendation engine using PySpark for curated product combinations via basket analysis. It involves fetching transaction data, preprocessing, building association models like FP-Growth and probability-based approaches, validating outputs, and scaling across departments to increase basket size, despite cold-start issues for new products.
E-commerce and Retail
Wayfair's Nightingale uses machine learning to optimize email cadence, predicting customer actions (purchase, unsubscribe, or nothing) after daily sales emails. Combining action probabilities with revenue and unsubscribe costs, it generates a net value score using over 500 customer features, improving AUC by 4%+ and reducing unsubscribe costs by 7% in A/B testing.
E-commerce and Retail
Wayfair's ABSA identifies aspects (e.g., comfort, delivery) in product reviews and determines sentiment, using a BERT-based self-attention model to surface products with positive attributes for long-tail pages. It achieved nearly a 2X improvement in macro-F1 score over the prior model.
E-commerce and Retail
Wayfair's NLP-powered virtual assistant resolves customer service issues 24/7 using text-based intent prediction. Built on the HaloAI platform, it identifies intents and provides automated resolutions, with plans to enhance accuracy and add multilingual capabilities.
E-commerce and Retail
DoorDash uses machine learning to manage supply and demand balance by predicting demand, optimizing supply, and dynamically pricing to ensure a balanced marketplace. They leverage historical data, real-time signals, and external factors to train models that predict demand and optimize supply. The models are integrated into the platform to dynamically adjust prices and ensure a balanced marketplace. Key challenges include handling data quality issues, dealing with cold starts, and ensuring fairness and transparency. DoorDash uses a combination of linear regression, gradient boosting, and neural networks to predict demand and optimize supply. They also use techniques like feature engineering, hyperparameter tuning, and model ensembling to improve model performance.
E-commerce and Retail
Stitch Fix developed a suite of algorithms called Style Explorer to predict which items will perform well and with whom, improving merchandise investments and reducing waste. Style Explorer maps items into a high-dimensional vector space using a neural network encoder, allowing comparison of items based on content rather than just interactions. It is trained using a skip-gram style objective to predict which items appear together in user sessions. Style Explorer's item-level performance predictions are leveraged by Algorithms partners and business partners across use cases like AI-assisted design, vendor evaluation, and the Elevate Grant & Mentorship Program.
E-commerce and Retail
Stitch Fix's approach involves mapping community inspiration images into a structured representation using computer vision, aggregating the item embeddings into a shared client-item latent space, scoring the relevance of each candidate item based on both distance in the latent space (context-relevance) and predicted client affinity (personalization), and selecting a diverse set of top-ranked items to recommend. This allows providing personalized recommendations that are also highly relevant to the specific query (e.g. a community inspiration image). The system leverages Stitch Fix's existing investments in latent style spaces and personalization models.
E-commerce and Retail
Accurate demand forecasting is critical for Ocado's large-scale ecommerce grocery ecosystem. They employ deep neural networks to ensure the best fit forecast models for their retail partners. Key challenges include balancing availability and minimizing stock waste, handling changing demand patterns across markets and geographies, and generating millions of forecasts daily.
E-commerce and Retail
Walmart developed a deep learning-based approach to improve product discoverability by correctly categorizing and shelving products in their e-commerce catalog. Key steps include identifying products with missing taxonomy hierarchy, preprocessing product data, training multi-class classification models like CNN-LSTM to predict the correct product type, using the model to assign the predicted taxonomy hierarchy, and executing rules to map to the right product category and family. The approach achieved 88% accuracy on the validation set and 80% on manual review, enabling Walmart to better categorize products and improve SEO.
E-commerce and Retail
Walmart used Fourier Transform to predict defrost in refrigeration cases, improving maintenance practices. The approach involves converting the case temperature signal to frequency domain, identifying the frequency component corresponding to defrost, filtering the frequency spectrum, and taking the inverse Fourier Transform to extract the defrost signal. The approach achieves over 80% precision and recall for cases with regular defrost patterns.
E-commerce and Retail
Mercado Libre developed global time series forecasting models to predict the next 12 weeks of sales and demand for each individual item sold by first party sellers in each country. The models use past sales history, questions asked, visits, price variations, and available stock as features. The sales model predicts a decrease in sales due to decreasing stock, while the demand model predicts sales recovery for items that run out of stock. The models were evaluated using MAE, MASE, MAPE, sMAPE metrics, with MAE chosen as most appropriate.
E-commerce and Retail
Wayfair built a central ads bidding platform to determine the appropriate amount to pay for millions of keywords and items in their catalog. Key aspects include using Single Bid Calculators to support different bidding approaches, providing observability into model decisions, allowing configurability of ROI and bid aggressiveness, enabling experimentation, and ensuring scalability. The platform leverages Google Cloud Dataproc, Aerospike, Python, Google BigQuery, and Java/K8s web services. Future plans include automated channel performance management, simulation capabilities, and multi-armed bandit approaches to rapidly tune algorithms.
E-commerce and Retail
Amazon developed a learning to rank (LTR) model to optimize package delivery by predicting the most likely location for a package to be delivered. The model uses a combination of features like GPS coordinates, package dimensions, and historical delivery data to rank potential delivery locations. This approach improved the accuracy of delivery location predictions, reducing the need for manual intervention and enhancing the overall customer experience.
E-commerce and Retail
Wayfair developed MARS, a transformer-based recommender system that uses sequential input data to adapt to changing customer preferences. MARS leverages self-attention to learn similarity between products and positional embeddings to capture the sequence of customer interactions. The model significantly improves recommendation accuracy by 67% and can learn transferable customer preferences across different product classes.
E-commerce and Retail
Amazon developed a learning to rank (LTR) model to optimize package delivery by predicting the most likely location for a package to be delivered. The model uses a combination of features like GPS coordinates, package dimensions, and historical delivery data to rank potential delivery locations. This approach improved the accuracy of delivery location predictions, reducing the need for manual intervention and enhancing the overall customer experience.
E-commerce and Retail
Zillow developed a scalable pipeline to generate embeddings from property listing descriptions using models like Word2Vec, GloVe, and BERT. These embeddings capture semantic similarity between listings and are integrated into Zillow's similar home recommendation system.
E-commerce and Retail
Mercado Libre faced challenges measuring the long-term impact of initiatives to retain valuable users. They developed a predictive model to identify at-risk users and send discounts to retain them. To estimate the true causal effect, they applied techniques like randomized controlled trials (RCTs), instrumental variables (IV), and observational studies with methods like inverse probability weighting (IPW), matching, and meta-learners. This allowed quantifying the incremental revenue and purchase frequency lift from their retention efforts over time. Key steps included identifying the causal graph, estimating incremental metrics, and validating estimator robustness.
E-commerce and Retail
Wayfair data scientists build scalable ML systems to programmatically optimize marketing decisions. Key approaches include general propensity models, specialized response/uplift models, and a multi-layer ML platform (WayLift) that leverages reinforcement learning to balance tradeoffs between exploiting the best ads and exploring alternative treatments.
E-commerce and Retail
Wayfair's SoV Optimization Engine maximizes the relevance of marketing messages shown to customers while meeting desired Share of Voice (SoV) targets. It uses a constrained optimization approach to find the optimal combination of customer-message pairs, represented as a 3D cube of Customer x Messages x Placements. Key inputs include customer-message relevance scores, placement value matrix, SoV targets, applicability matrix, and other constraints. The optimization problem is defined using Gurobi and solved in batches.
E-commerce and Retail
Wayfair has developed a scalable algorithmic decision-making platform called WayLift that uses contextual bandit algorithms to optimize millions of daily customer-level marketing and advertising decisions. The contextual bandit model consists of a reward estimator and an exploration algorithm to balance exploration and exploitation.
E-commerce and Retail
Wayfair uses explainable AI techniques like permutation importance, LIME, and SHAP to interpret the decisions of their fraud detection models. This supports auditing for bias and assists human reviewers in investigating flagged orders.
E-commerce and Retail
OLX uses triplet loss to detect fraudulent images in online advertising by training a Siamese network to generate embeddings that capture image similarity.
E-commerce and Retail
ViCs is an image-based deep learning model that uses a Siamese network with triplet loss to understand product compatibility across classes, generating stylistically similar recommendations.
E-commerce and Retail
Stitch Fix uses BERT to extract information from client Fix Notes, fine-tuning the model to feed dynamic word embeddings into their styling algorithm.
E-commerce and Retail
Instacart uses datashader to visualize massive GPS data from grocery delivery shoppers, aiding in logistics optimization and operational improvements.
Travel, E-commerce and Retail
This AI system is designed to enhance the user experience on Airbnb by prioritizing and highlighting home attributes that align with guest interests. By analyzing user interactions, reviews, and booking patterns, the system identifies which features of a listing (such as amenities, location, and property type) are most appealing to potential guests. This prioritization helps in tailoring search results and recommendations, making it easier for users to find homes that meet their preferences.
Travel, E-commerce and Retail
This AI system developed by Airbnb focuses on improving the diversity of search result rankings to enhance user engagement and satisfaction. By incorporating diverse results in search rankings, the system ensures that users are exposed to a wider range of listings, thereby increasing the likelihood of finding options that meet their unique preferences. The system uses advanced machine learning techniques to balance relevance and diversity in search results, optimizing the overall user experience.
Travel, E-commerce and Retail
This AI system developed by Airbnb focuses on categorizing listings based on various attributes to enhance user search and browsing experiences. By integrating machine learning with a human-in-the-loop approach, the system ensures that the categorization is accurate, relevant, and continuously improving. The ML models analyze large amounts of data to identify patterns and suggest categories, while human experts validate and refine these suggestions to maintain high quality and relevance.
Travel, E-commerce and Retail
This AI system developed by Expedia leverages synthetic search data to improve the accuracy and reliability of flight price forecasting. Synthetic data, which is artificially generated rather than collected from real-world events, allows for more extensive and diverse training datasets. This helps the system to better predict future prices by accounting for a wider range of scenarios and reducing the biases present in real-world data.
Travel, E-commerce and Retail
The Customer Lifetime Value (CLV) Prediction Model developed by Expedia Group aims to forecast the long-term value of customers. This model helps in identifying high-value customers, optimizing marketing strategies, and enhancing overall customer relationship management. The CLV model uses historical booking data, customer interactions, and various other features to predict the future revenue a customer is likely to generate over their lifetime with Expedia.
Travel, E-commerce and Retail
This AI system developed by Expedia focuses on generating diverse travel recommendations to enhance user engagement and satisfaction. The system leverages machine learning techniques to analyze user preferences, past interactions, and various attributes of travel destinations to provide a wide range of personalized travel options. The goal is to inspire users to explore different locations and experiences, thereby increasing the likelihood of bookings and user retention.
Travel, E-commerce and Retail
This AI system developed by Expedia Group focuses on increasing traveler engagement by sending relevant price alerts. The system analyzes flight and accommodation prices, along with user preferences and behavior, to provide timely notifications about price drops and deals. This approach helps users find the best prices for their trips, enhancing their booking experience and increasing the likelihood of conversions.
Travel, E-commerce and Retail
This section typically provides an overview of the AI system, including its purpose, functionality, and the problem it aims to solve.
Travel, E-commerce and Retail
This article discusses the balance between exploration (showing new accommodations) and exploitation (showing known well-performing accommodations) in trivago’s ranking model. The goal is to optimize the accommodation ranking to benefit users, advertisers, and trivago by finding the optimal order to display accommodations.
Travel, E-commerce and Retail
Find a detailed explanation of what the AI system does. This section should describe its purpose, functionality, and the problem it aims to solve.
Travel, E-commerce and Retail
Find a section that explains what the AI system does. This should include its purpose, the problem it addresses, and a high-level overview of its functionality.
Travel, E-commerce and Retail
Find a detailed explanation of the AI system, describing its purpose, how it works, and its main features. This section usually provides an overview of the system's functionality.
Travel, E-commerce and Retail
The personalized ranking model for lodging is an AI-driven system designed to improve the user experience on lodging platforms by tailoring search results according to individual user preferences. This model leverages various data points, including user behavior, past booking history, and demographic information, to rank lodging options in a manner most likely to appeal to the user. The primary goal is to increase conversion rates and user satisfaction by presenting the most relevant and personalized lodging options at the top of search results.
Travel, E-commerce and Retail
The AI system aims to enhance Airbnb's search ranking by improving their deep learning models. The improvements include developing more sophisticated neural network architectures, addressing positional bias in rankings, and solving the cold start problem for new listings (ar5iv) (DeepAI).
Travel, E-commerce and Retail
Airbnb Experiences are unique, handcrafted activities designed and led by local hosts. Since its launch in November 2016, the service has expanded to over 1,000 destinations worldwide. To support this growth and ensure that users can easily find relevant experiences, Airbnb developed a machine learning-powered search ranking system. This system personalizes search results based on various factors such as user preferences, experience quality, and current trends. Over two years, Airbnb ran numerous experiments to fine-tune the algorithm, significantly improving booking rates and user satisfaction. The system includes features like real-time model scoring, personalized navigation, and ranking dashboards for insights into predictions and optimization (mlconf) (huyenchip).
1 use case
The system leverages causal inference methods to understand the impact of product changes, content localization, and other business decisions on user engagement and satisfaction. It helps Netflix make data-driven decisions by measuring the true causal effects rather than mere correlations.
1 use case
A multi-component machine learning system designed to extract metadata from text-heavy user-uploaded documents. The system uses keyphrase and entity extraction to provide a semantic understanding of documents. Keyphrases represent major themes, while entities are proper nouns such as people, places, and organizations. This combination enriches downstream discovery models and helps categorize documents more accurately.
17 use cases
Media and streaming
The article discusses the ML models that Spotify uses to recommend music and podcasts on the Home page. The models select the best content (candidate generation) and rank it for each listener. Specific models include the Podcast Model, Shortcuts Model, and Playlists Model, each designed for different content types and user interactions.
Media and streaming
The AI system combines machine learning algorithms with reader feedback to recommend articles to NYTimes readers. The recommendation engine uses data on reader behavior and preferences to personalize the reading experience and improve user engagement.
Media and streaming
This system transforms from collaborative filtering to Transformer-based models for sequential recommendations, creating embeddings that capture semantic information about documents and users. It addresses the challenge of delivering personalized content to users in real-time using embeddings for efficient similarity searches.
Media and streaming
The article discusses various machine learning models employed by Spotify to recommend personalized content on the Home page. These models include the Podcast Model, Shortcuts Model, and Playlists Model, each designed to cater to different user interactions and preferences, enhancing the user experience by tailoring recommendations.
Media and streaming
The system leverages computer vision models to categorize documents based on visual features, aiming to distinguish between types such as text-heavy documents, comics, and spreadsheets. It combines semantic understanding with user behavior to create a multi-component machine learning system that enhances document discovery.
Media and streaming
This system uses deep learning algorithms to analyze content and context, delivering more relevant advertisements. By understanding the semantic meaning of the content, it can match ads more effectively, improving user engagement and ad performance.
Media and streaming
This system uses machine learning algorithms to analyze user-uploaded documents, determining their category (e.g., educational, professional) to enhance targeted advertising. It relies on contextual and content analysis to deliver more relevant advertisements, improving user engagement and ad performance.
Media and streaming
This system was developed to provide personalized shortcuts on the Spotify Home page, allowing users to quickly access their favorite content. It uses heuristics and machine learning models to predict and display the top entities a user is most likely to play next.
Media and streaming
The system uses cross-lingual transfer learning to classify content into appropriate categories, leveraging models trained on multilingual data. This approach helps DailyMotion manage and organize vast amounts of content uploaded in various languages, improving search and recommendation accuracy.
Media and streaming
The system utilizes machine learning algorithms to enhance the relevance and diversity of video recommendations. By balancing user preferences with diverse content, the system aims to keep users engaged and satisfied with varied video suggestions.
Media and streaming
The system automates bid adjustments and budget allocations across various marketing campaigns, utilizing machine learning to enhance decision-making and optimize spending efficiency based on historical data and predictive analytics.
Media and streaming
Utilizes machine learning to optimize ad spending across various platforms. MSOE analyzes performance data to adjust marketing tactics dynamically, focusing on maximizing ROI by identifying the most effective channels and strategies.
Media and streaming
This section would detail how Lyft uses machine learning and automation technologies to optimize its marketing efforts, improving targeting, and personalization to enhance user engagement and cost-efficiency.
Social platforms
The system is designed to predict advertiser churn within the next 14 days, classifying accounts into high, medium, and low churn risk categories. It uses over 200 features and the SHAP library to estimate feature contributions to the model’s output. The Sales team uses these predictions to mitigate churn, resulting in a 24% reduction in churn rate for high-tier pods in treatment experiments.
Social platforms
Call classification systems using machine learning (ML) and natural language processing (NLP) aim to automate the categorization of incoming customer calls. These systems analyze call logs and identify the reasons behind customer calls, which helps in routing them to the appropriate department or providing automated responses. The primary goal is to enhance operational efficiency, reduce handling time, and improve customer satisfaction by quickly addressing their needs.
Social platforms
This system is designed to identify and mitigate viral spam content on LinkedIn. It employs AI models to detect and prevent spam, ensuring a safe and professional environment for users. The models are categorized into proactive defenses (early detection) and reactive defenses (monitoring and action based on engagement).
Social platforms
Yelp's system involves transforming various types of content (reviews, photos, and business information) into embeddings, which are dense vector representations capturing semantic meaning. These embeddings facilitate numerous downstream applications like search, recommendation, and personalization by encoding rich contextual information about Yelp's data.
12 use cases
A sophisticated system developed by Gojek to enhance cross-selling strategies by accurately matching products with users. This AI-driven algorithm analyzes user behavior to predict which products they might need next, improving product discovery and user satisfaction. It uses advanced machine learning techniques to process vast amounts of user data.
The article describes the optimization of DoorDash's search functionality using advanced machine learning techniques to better understand user intent, enhancing the accuracy of search results and improving customer satisfaction.
The article would describe how the AI system uses deep learning algorithms to accurately detect and classify stop signs and traffic signals from visual data collected via vehicle-mounted cameras, helping improve navigation and safety features.
Gojek's AI system that tailors search results for individual users to enhance user satisfaction and conversion rates. This AI system leverages learning to rank algorithms to personalize restaurant listings in GoFood searches, improving relevance based on individual preferences and past interactions.
The system likely uses advanced machine learning techniques to predict customer behavior regarding items with little to no historical data, improving product discovery and recommendation accuracy.
The description might discuss how the system integrates open-source map data with real-time data from various sources, using machine learning algorithms to continuously update and refine map accuracy.
The system utilizes advanced machine learning models to understand user intent and optimize search relevancy by analyzing actions like clicks and orders. This improves user experience by aligning search results more closely with user expectations.
Gojek’s automated forecasting tool designed to incorporate external variables into predictive models using machine learning techniques. GO-FORECAST simplifies the entire machine learning pipeline, allowing Gojek's team members to tune, train, and predict with just a few clicks. The system utilizes XGBoost, a powerful algorithm, to enhance forecasting accuracy across various datasets.
This system uses advanced statistical models or machine learning algorithms to predict long-term trends, including demand, supply dynamics, and market changes.
The system uses data analytics and machine learning to identify patterns and anomalies that suggest fraudulent activities, enhancing security and trust on the Lyft platform.
This system transitions from shallow learning to deep learning techniques in fraud detection, improving accuracy and the ability to detect complex fraudulent patterns on Lyft’s platform.
The system leverages machine learning to tailor marketing efforts to individual user preferences, improving engagement and conversion rates.
2 use cases
Zillow developed models to detect and localize windows, doors, and openings in indoor 360-degree panoramas using bounding boxes, trained on a custom dataset to support floor plan generation.
Zillow's Guided Search suggests personalized search refinements using machine learning to rank relevant filters based on user context, helping customers narrow their home searches.
2 use cases
The description usually follows the introduction of the AI system. It should explain what the AI system does, its purpose (in this case, related to course recommendations), and possibly its impact or significance within LinkedIn's platform. This system likely recommends relevant courses to users based on their profiles and activity.
The description usually follows the introduction and explains what the AI system does in the context of course recommendations. It should cover its purpose, functionality, and possibly its impact or significance within LinkedIn's platform. This system likely provides personalized course suggestions to enhance user learning.
1 use case
Purpose and Functionality: Typically found in the introductory paragraphs or sections, the description will explain what the AI system does in the context of image detection as a service. This system likely identifies objects or patterns in images for various applications.
7 use cases
The HR Digital Assistant leverages large language models to assist HR professionals with everyday tasks. It can provide real-time insights, help with writing job descriptions and offer letters, create custom analytics dashboards, and offer guidance on employment law compliance through a conversational interface.
The HR Automation Assistant utilizes generative AI to enable users to create workflow automations without needing expert coders. By interpreting plain English instructions, it can generate the necessary code to automate repetitive HR tasks. For instance, it can automatically approve timesheets based on predefined criteria, ensuring efficient and accurate task management.
The AI-Powered Onboarding Assistant enhances the new-hire onboarding experience by translating and summarizing HR policies, generating personalized onboarding journeys, guiding benefits selections, and providing HR practitioners with detailed analytics and insights. This assistant aims to streamline the onboarding process, making it seamless and engaging for new employees while reducing the workload for existing staff.
The AI Virtual Mentor provides personalized training assistance by leveraging access to all enterprise files and policies. This system can offer quick answers to employee queries, guide them through training programs, and ensure they receive relevant and timely information, enhancing their learning experience.
The AI-Powered Payroll Optimizer leverages advanced AI techniques to detect payroll calculation and input errors, fraud, legal risks, and exceptions. It analyzes changes in payment patterns, user behavior, and pay rules to identify and rectify issues. The system also incorporates Robotic Process Automation (RPA) to streamline payroll processing and ensure accuracy.
The AI-Driven Employee Engagement and Wellbeing Monitor uses AI to analyze extensive employee data, such as surveys, emails, and performance metrics, to identify trends and potential issues. This system helps HR pinpoint employees at risk of burnout and provides personalized support through AI-powered chatbots. It also integrates into company platforms to create a data-driven Digital Employee Experience (DEX), which monitors engagement, learning progress, and work performance to make informed HR decisions.
Generative AI analyzes the goals set for each employee and compares their actual performance against these goals. This analysis helps identify areas where employees excel and where they encounter difficulties, enabling managers to provide targeted feedback and develop individual development plans.
Social Media
1 use case
Predicting Ad Request Value: Machine Learning Insights from Twitter
The description usually follows the introduction of the AI system. It should explain what the AI system does, its purpose (in this case, predicting ad request value), and possibly its impact or significance within Twitter's platform. This system likely uses machine learning to estimate the value of ad requests for better ad placement.
Automation feasibility: Yes