# recommendation systems Machine-learning systems that select and rank items to provide users with personalized recommendations from a large collection. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## ML based ranking using Nrtsearch DevFeed: [ML based ranking using Nrtsearch]() Original publisher: [Read original article]() Author: Mohammad Mohtasham (Software Engineer); Tao Yu (Software Engineer) Published: 2026-09-16T00:00:00Z Content type: article Language: en Sources: [Yelp]() Topics: [Inference](), [Scalability](), [Latency](), [recommendation systems](), [Open Source]() Tags: [bridge](), [inference](), [machine-learning](), [overhead](), [performance](), [pipeline](), [scalability](), [service]() ### AI overview Yelp's Nrtsearch Inference Plugin embeds machine-learning ranking directly in the search layer. The article explains the scoring workflow, including model configuration, feature extraction, candidate ranking, and application-specific business logic. It describes how co-locating feature storage and inference reduces network transfer, serialization overhead, and latency compared with a standalone inference service. ### Source excerpt We've extended Nrtsearch with the Inference Plugin, which embeds ML-based ranking directly in the search layer -- eliminating the need for a standalone scoring service. We use Nrtsearch (read more information on the blog post), a Lucene-based open-source search engine built by Yelp, to power a variety of applications such as business search, reviews search, ad delivery and photo search. In this blog post, we give a high-level overview of the Machine Learning (ML) based scoring workflow in Nrtsearch. We'll show how ML models are configured and loaded, and how different applications use custom business logic to develop, test, and... ## 🍔🧠 Pinterest's Fix for the Hardest Problem in ML Infra DevFeed: [🍔🧠 Pinterest's Fix for the Hardest Problem in ML Infra]() Original publisher: [Read original article]() Author: Alexandre Zajac Published: 2026-09-14T15:31:30Z Content type: article Language: en Sources: [Hungry Minds]() Topics: [recommendation systems](), [Feature Engineering](), [Inference](), [Streaming](), [configuration](), [Framework](), [data](), [Low Latency]() Tags: [batch](), [configuration](), [data](), [feature-engineering](), [low-latency](), [ml](), [pinterest](), [ranking](), [recommendation-systems](), [streaming]() ### AI overview Pinterest redesigned its user-sequence platform for ranking, retrieval, and recommendation systems by defining signals once and instantiating them consistently across streaming, batch, and serving workloads. The approach uses Python configuration with validated schemas, a shared execution engine, cooperating streaming and batch paths, and columnar time-partitioned storage to improve freshness, completeness, consistency, and operational efficiency. ### Source excerpt PLUS: OpenAI agents beat math 🧮, Test techniques for agents ⚡, Postgres survival guide 📖 ## How LinkedIn Trains AI Job Search 8x Faster with Multi-Teacher Distillation DevFeed: [How LinkedIn Trains AI Job Search 8x Faster with Multi-Teacher Distillation]() Original publisher: [Read original article]() Author: Claudio Masolo Published: 2026-09-11T10:00:00Z Content type: news Language: en Sources: [InfoQ]() Topics: [recommendation systems](), [Large Language Model](), [Language models](), [sglang](), [Latency]() Tags: [agentic-ai-architecture](), [ai](), [ai-ml-data-engineering](), [artificial-intelligence](), [large-language-models](), [latency](), [liger](), [linkedin](), [linkedin-ai-multi-teacher](), [llms](), [ml-data-engineering](), [news](), [search](), [sglang](), [training]() ### AI overview LinkedIn describes a multi-teacher distillation pipeline for AI-powered job search that trains a 0.6B-parameter ranking model. The article focuses on SGLang-based teacher serving, online and offline distillation, and training optimizations reported to produce roughly an eightfold speedup. ### Source excerpt LinkedIn has published details of the training infrastructure behind its AI-powered job search, describing a multi-teacher distillation pipeline that compresses knowledge from large teacher models into a compact 0.6B-parameter ranking model. By Claudio Masolo ## MAPS: Netflix's Multimodal Asset Personalization at Scale DevFeed: [MAPS: Netflix's Multimodal Asset Personalization at Scale]() Original publisher: [Read original article]() Author: Netflix Technology Blog Published: 2026-08-28T16:01:02Z Content type: article Language: en Sources: [Netflix](), [Netflix TechBlog - Medium]() Topics: [Netflix](), [personalization](), [Embeddings](), [multimodal](), [recommendation systems](), [data](), [A/B Testing]() Tags: [data](), [embedding](), [embeddings](), [machine-learning](), [model](), [multimodal](), [netflix](), [personalization](), [production](), [testing]() ### AI overview Netflix describes MAPS, a multimodal asset-personalization system that uses embeddings from artwork and video previews to address cold-start problems and personalize recommendations soon after a title launches. ### Source excerpt By Emma Yanyang Kong, Aditya Deshpande, Asad Abbasi, Bowei Yan, David Fagnan, Ashish Rastogi, Dhaval Patel, Ray Zhang Introduction The Netflix experience is a journey of discovery. Every visual cue, from the artwork on a title to the video previews that autoplay while you browse, is there to connect you with a story you will love. We call these visual cues assets, and choosing the right one for each member is a personalization problem of its own. But which image or video preview of Squid Game should we show you? And what do we do right after a title launches, when there's far too little interaction data to know which asset we should recommend to each member? For years, our models answered the first question well and the second poorly. They learned which assets members interacted with, but treated every asset as an opaque ID, blind to what was actually in the artwork or video preview. Right after a title launched, its assets had no history, so we dialed up exploration on its assets to gather interaction data, and otherwise fell back to popularity heuristics that ignore your taste. Only once enough interactions had piled up could personalization take over. This is the classic cold-start problem. This post shares how multimodal embeddings let our models see and hear the assets they recommend, so personalization can kick in far sooner, close to a title's launch. Because a new asset arrives with its embedding the model already understands, that embedding carries member taste signals from related assets immediately. Consequently, the model needs far less interaction history before it can personalize. We cover three production systems, artwork personalization, query-aware artwork ranking, and video preview personalization, plus a cheap trick for choosing new embeddings before committing to full end-to-end integration and A/B testing. Artwork Personalization A single image is often a member's first touchpoint with a title, so we create a diverse set of artworks for each tit ## Scaling Conditional Learned Retrieval for Pinterest Home Feed DevFeed: [Scaling Conditional Learned Retrieval for Pinterest Home Feed]() Original publisher: [Read original article]() Author: Pinterest Engineering Published: 2026-08-26T14:01:05Z Content type: article Language: en Sources: [Pinterest Engineering Blog - Medium]() Topics: [recommendation systems](), [Embeddings](), [Machine Learning & Artificial Intelligence]() Tags: [blog](), [embedding](), [eng-culture](), [engineering](), [infrastructure](), [pinner-experience](), [pinterest](), [ranking](), [retrieval]() ### AI overview Pinterest describes scaling Conditional Learned Retrieval for home-feed candidate generation. The approach extends two-tower retrieval with explicit contexts that produce multiple user embeddings for different interests and retrieval use cases. ### Source excerpt Devin Kreuzer | Sr. Machine Learning Engineer; Yichi Wang | Machine Learning Engineer I; Sujan Reddy Ale | Machine Learning Engineer I; Zelun Wang | Sr. Machine Learning Engineer; Hongtao Lin | Sr. Machine Learning Engineer; Piyush Maheshwari | Staff Machine Learning Engineer Pinterest home feed candidate generation is a large-scale User-to-Pin retrieval problem. A common approach is a two-tower model: a user tower encodes the user, an item tower encodes candidate Pins, and approximate nearest neighbor search retrieves Pins close to the user embedding. But Pinterest users often have multiple intentions at once -- planning a renovation, saving recipes, exploring fashion, or organizing travel ideas. A single retrieval embedding can struggle to capture this diversity. Conditional Learned Retrieval, or CLR, extends the two-tower setup by conditioning the user tower on an explicit retrieval context. Instead of producing only one user embedding, CLR can generate condition-aware embeddings that reflect different aspects of a user's interests while still grounding retrieval in the user's overall behavior. Prior Pinterest work studied this formulation in two settings. The RecSys'24 paper: Bootstrapping Conditional Retrieval for User-to-Item Recommendations described how to bootstrap conditional retrieval by constructing training data for (user, condition) -> item retrieval from existing user-item and condition signals, and applied it to interest-based notifications. The KDD'25 paper: Synergizing Implicit and Explicit User Interests: A Multi-Embedding Retrieval Framework at Pinterest placed Conditional Retrieval within a broader multi-embedding retrieval framework for home feed, where explicit interest conditions complement implicit interests extracted from user behavior. In this blog, we describe how CLR evolved from early interest-conditioned retrieval into a broader retrieval system for Pinterest home feed. We focus on three areas: expanding CLR to support more retrieval us ## MTIA 300: Meta's First Training Chip with Built-in NICs and Communication-Offloading Engines DevFeed: [MTIA 300: Meta's First Training Chip with Built-in NICs and Communication-Offloading Engines]() Original publisher: [Read original article]() Author: Rajiv Krishnamurthy; Wes Bland Published: 2026-08-24T17:45:52Z Content type: article Language: en Sources: [Engineering at Meta]() Topics: [recommendation systems](), [Language models]() Tags: [chip-design](), [communication](), [data-infrastructure](), [devinfra](), [gpu](), [hardware](), [inference](), [infrastructure](), [meta](), [networking-traffic](), [performance](), [production-engineering](), [training]() ### AI overview Meta describes MTIA 300, an in-house accelerator for training ranking and recommendation models, with built-in network chiplets and a co-designed HCCL communication library. The design targets communication-heavy distributed training by integrating RDMA NICs into the chip package and offloading communication work. ### Source excerpt MTIA 300 is the first of Meta's family of in-house training and inference accelerators optimized for training ranking and recommendation models. We're sharing how MTIA 300's built-in NIC chiplets allow it to meet the communication needs associated with training recommendation models with superior performance over general-purpose GPUs. By co-designing MTIA's communication library, HCCL, alongside the [...] Read More... The post MTIA 300: Meta's First Training Chip with Built-in NICs and Communication-Offloading Engines appeared first on Engineering at Meta. ## How Generative Recommenders Are Redefining RecSys at Scale DevFeed: [How Generative Recommenders Are Redefining RecSys at Scale]() Original publisher: [Read original article]() Author: Elizabeth Goodman Published: 2026-08-20T16:00:00Z Content type: article Language: en Sources: [NVIDIA Developer](), [NVIDIA Technical Blog]() Topics: [recommendation systems](), [AI, ML & Data Engineering](), [Training AI Models](), [Language models]() Tags: [agentic-ai-generative-ai](), [ai-inference](), [data-center-cloud](), [developer-tools-techniques](), [embedding](), [featured](), [generative](), [gpu](), [inference](), [latency](), [llms](), [machine-learning](), [machine-learning-artificial-intelligence](), [recommenders-personalization](), [scale]() ### AI overview The article examines the shift toward generative recommender systems and the challenges of training and serving them at large scale. ### Source excerpt Recommender systems (RecSys) are one of the most ubiquitous machine learning problems in the consumer internet industry yet notoriously difficult to train and... ## How Keras 3 Helped Modernise Expedia Group's Lodging Ranking Stack DevFeed: [How Keras 3 Helped Modernise Expedia Group's Lodging Ranking Stack]() Original publisher: [Read original article]() Author: Conor Worthington Published: 2026-08-11T11:01:02Z Content type: article Language: en Sources: [Expedia]() Topics: [Keras](), [recommendation systems](), [Latency](), [Inference Performance](), [Machine learning]() Tags: [data-science](), [features](), [framework](), [improvements](), [inference](), [keras](), [latency](), [llms](), [machine-learning](), [mixture-of-experts](), [model-training](), [neural-networks](), [ranking](), [recommendation-system]() ### AI overview Expedia Group describes modernising its lodging-ranking stack around Keras 3. The changes included rewriting parts of its pipelines, making model training 30% faster and reducing P99 inference latency by two-thirds. The article discusses limitations of the previous Keras 2, eager-training and combined-serving setup, along with the use of newer APIs and accelerator-oriented optimisations. ### Source excerpt Expedia Group Technology -- DataWhat happened when we treated a framework migration as an architecture modernisation -- and cut P99 inference latency by two-thirdsSt Paul's and millennium bridge, London Expedia Group™ has always been a market leader in providing personalised search experiences for travellers. As our ranking models evolved, we saw an opportunity not just to migrate to Keras 3, but to modernise the broader stack around it so we can better serve travellers. This led us to rewrite key parts of our pipelines that made model training 30% faster and cut P99 inference latency by two-thirds. Our main focus in this blog is to discuss our improvements to lodging ranking -- our service which handles users' search requests and returns a personalised property ranking on top of a lightweight candidate generator. This service is frequently retrained to improve customer experience as we get new signals over time. Example of personalised lodging ranking on Expedia search result page for a London search As such, our velocity to make changes needs to be fast, but more importantly we need to be able to build models which are state-of-the-art, enabling customers to easily find and book the most relevant property for their trip. The problem with Keras 2, eager training and combined serving The challenge was not that the old stack was broken. It was that it had gradually become a limiting factor. On the training side, we were carrying dependencies on older components and missing out on modern Keras APIs, new optimisers, new layers and cleaner support for accelerator-oriented training. On the serving side, we had a path that was simple and operationally familiar, but not especially well suited to aggressively optimised inference for heavier ranking architectures. This blockage on utilising new layers and optimisers has become more pronounced as LLMs see blistering amounts of innovation. Naturally, these new features are only available in Keras 3 or as separate backends from ou ## From User Sequences to Scaling Laws: A Multi-Stage Architecture for Meta's Ads Ranking DevFeed: [From User Sequences to Scaling Laws: A Multi-Stage Architecture for Meta's Ads Ranking]() Original publisher: [Read original article]() Author: Steven De Gryze; Parshva Doshi; Sean O'Byrne; Arnold Overwijk; Dinesh Ramasamy; Lee Xiong Published: 2026-08-05T19:20:20Z Content type: article Language: en Sources: [Engineering at Meta](), [Meta ML Applications]() Topics: [recommendation systems](), [Temporal data]() Tags: [ads](), [architecture](), [data-infrastructure](), [meta](), [ml-applications](), [production-engineering](), [recommendation-systems](), [scaling-laws](), [tokenization]() ### AI overview Meta describes a multi-stage sequence-model architecture for ads ranking that separates offline user modeling from lightweight online ranking. It also uses dense tokenization and target-aware attention to learn feature interactions, with reported conversion and ad-click lifts across Instagram and Facebook. ### Source excerpt Every day, Meta's recommendation platforms handle billions of user interactions, generating rich temporal signals that capture individual preferences and intent across products, ads, and content. In our 2024 post on sequence learning for ads recommendations, we showed how modeling the order and timing of user actions (rather than relying on static, manually engineered sparse features) [...] Read More... The post From User Sequences to Scaling Laws: A Multi-Stage Architecture for Meta's Ads Ranking appeared first on Engineering at Meta. ## GEM Training: How Meta Doubled the Efficiency of Its LLM-Scale Ads Foundation Model DevFeed: [GEM Training: How Meta Doubled the Efficiency of Its LLM-Scale Ads Foundation Model]() Original publisher: [Read original article]() Author: Darren Liu; Huayu Li; Raghav Boinepalli; Yuzhen Huang; Jackie (Jiaqi) Xu; Richard Qiu; Chunzhi Yang; Rich Zhu; Dev (Devashish) Shankar; Huaqing Xiong Published: 2026-08-03T18:00:17Z Content type: article Language: en Sources: [Engineering at Meta](), [Meta AI Research](), [Meta ML Applications]() Topics: [recommendation systems](), [networking]() Tags: [ads](), [ai-infrastructure](), [ai-research](), [data-infrastructure](), [gpu](), [hardware](), [kernels](), [llm](), [meta](), [ml-applications](), [networking](), [recommendation-systems](), [training]() ### AI overview Meta describes training its GEM ads recommendation foundation model at LLM scale. The article covers recommendation-specific kernels, ultra-low-precision training, and topology-aware parallelism that doubled end-to-end training efficiency to 20-25% MFU while increasing training FLOPs fourfold. ### Source excerpt Meta's Generative Ads Recommendation Model (GEM), the foundation model behind ads recommendations across Instagram and Facebook, now trains at LLM scale on several thousand of the latest-generation GPUs. This post goes into the details on how we achieved: doubling end-to-end (E2E) training efficiency to 20-25% Model FLOPs Utilization (MFU) while scaling training FLOPs 4x in [...] Read More... The post GEM Training: How Meta Doubled the Efficiency of Its LLM-Scale Ads Foundation Model appeared first on Engineering at Meta. ## GenRec: Towards LLM-Native Recommendation at Netflix DevFeed: [GenRec: Towards LLM-Native Recommendation at Netflix]() Original publisher: [Read original article]() Author: Netflix Technology Blog Published: 2026-07-30T20:10:15Z Content type: article Language: en Sources: [Netflix](), [Netflix TechBlog - Medium]() Topics: [Netflix](), [recommendation systems](), [Large Language Model](), [LLM Techniques](), [personalization](), [vllm]() Tags: [data](), [feature-engineering](), [genai](), [large-language-models](), [large-language-models-llms](), [llm](), [llms](), [netflix](), [personalization](), [ranking](), [recommendation-system](), [vllm]() ### AI overview Netflix presents GenRec, an LLM-backed recommendation ranker trained on Netflix-specific data and objectives. It converts user histories, item metadata, and context into text, uses a catalog-aware scoring head, aligns recommendations with long-term member value and business goals, and runs in prefill-only mode on Netflix's LLM serving stack. In a large-scale A/B test, GenRec improved short- and long-term online metrics while using fewer labeled examples and input signals than a mature production ranker. ### Source excerpt Authors: Ying Li, Arjun Rao, Shradha Sehgal Introduction Recommendations sit at the heart of the Netflix experience. Our current production models rely on thousands of hand-crafted features over users, items, and interactions, along with specialized architectures for sequence modeling, feature interactions, and multi-task objectives. This stack has evolved over many years to support diverse content types (movies, series, games, live, podcasts) and product surfaces, but its complexity makes it costly to onboard new use cases: adding a content type or surface can require significant feature engineering, architecture change, infrastructure work, and experimentation. At the same time, large language models (LLMs) are changing how we think about recommendation, as shown by recent work such as PLUM, GLIDE, and OneRec-Think. Their broad world knowledge and strong language understanding make it possible to represent user histories and item metadata directly as text, capture rich relationships in a shared semantic space, and steer recommendations via natural-language prompts. However, off-the-shelf LLMs are still far from production-ready recommenders: they often over-recommend globally popular content, hallucinate out-of-catalog items, ignore business constraints, and provide only limited personalization. To address this, we built GenRec, an LLM-backed recommendation ranker that post-trains an internal foundation LLM on Netflix-specific data and objectives. GenRec shows that an LLM-based ranker can match or exceed a mature production system while relying on far fewer labeled examples and input signals. Figure 1: GenRec pipeline. Raw logs of user history, item metadata, and context are transformed via context engineering into natural-language prompts and fed into the GenRec, which runs on vLLM in prefill-only mode and outputs scores for each catalog item, yielding a recommendation ranking. At a high level, GenRec: Verbalizes user histories, item metadata, and context as text ## Pinner Progression: Better Use-Case Representation Driving Weekly Active User Growth at Pinterest DevFeed: [Pinner Progression: Better Use-Case Representation Driving Weekly Active User Growth at Pinterest]() Original publisher: [Read original article]() Author: Pinterest Engineering Published: 2026-07-27T16:01:02Z Content type: article Language: en Sources: [Pinterest Engineering Blog - Medium]() Topics: [recommendation systems](), [Data Science](), [data]() Tags: [engineering](), [growth](), [interest-exploration](), [metrics](), [personalization](), [pinterest](), [recommendation-system](), [retention](), [retrieval](), [understanding-user]() ### AI overview Pinterest introduces Pinner Progression, a recommendation-system program that uses persistent User Interest Clusters to improve discovery and make retention a first-class objective alongside engagement. ### Source excerpt Part 1 of 2 Authors Personalization (Homefeed): Yuke Yan, Chuxi Wang, Andreanne Lemay, Olafur Gudmundsson, Anna Kiyantseva, Krystal Benitez, Jongho Kim, Jiacong He, Rahul Goutam, James Li, Dylan Wang User Understanding: Simin Li, Sufyan Suliman, Yingjian Ding, Hongbo Deng Data Science: Armando Ordorica, Yan Chen, Ellie Zhang, Karim Wahba Introduction Pinterest's mission is to help people discover the inspiration to create a life they love. Our recommendation system serves hundreds of millions of users, surfacing billions of Pins across interests ranging from home renovation to meal planning to wedding decor. The home feed , where much of that discovery happens, is powered by a multi-stage pipeline spanning retrieval, lightweight scoring, ranking, and re-ranking [1][2][3]. Most of our prior work on this pipeline has been optimized for engagement: clicks, saves, downloads, closeups. These are strong signals of immediate relevance, and optimizing for them has driven significant gains across the system [4][5]. The problem is that engagement and retention are different things. A user can save ten sourdough recipes today and churn next month anyway. All we did was feed them more of what they already liked: we never helped them find something new for next time. This post is one of two that introduces Pinner Progression, a program that reframes the home feed recommendation system around retention as a first-class objective. Our core insight is that by augmenting sequential, action-by-action user understanding with holistic, persistent use-case representation, we can reliably anticipate the user's next moves and start to serve recommendations that ignite their serendipitous discovery. In this post, we introduce the key use-case representation signal: User Interest Clusters (UICs) -- and describe its construction, integration into the recommendation stack, and impact on engagement on retention metrics. A follow-up will cover how we predict unseen UICs and conduct systematic us ## Personalizing Airbnb search by learning from the guest journey DevFeed: [Personalizing Airbnb search by learning from the guest journey]() Original publisher: [Read original article]() Author: Daochen Zha Published: 2026-07-21T17:01:04Z Content type: article Language: en Sources: [The Airbnb Tech Blog - Medium]() Topics: [Transformer](), [recommendation systems](), [Temporal data](), [Data Science]() Tags: [ai](), [blog](), [conversion](), [engineering](), [machine-learning](), [model](), [recommendation-system](), [research](), [scale](), [search](), [technology](), [training]() ### AI overview Airbnb describes a Transformer-based sequence model for personalizing search by encoding years of guest behavior, including listing views, bookings, reviews, and cancellations. The system learns richer representations of guest preferences to improve listing relevance and booking conversion while addressing very long, noisy event sequences and the cost of training on hundreds of millions of search-label pairs. ### Source excerpt How we built a Transformer-based sequence model that encodes years of guest behavior to surface the right listings at the right time. By: Daochen Zha, Chun How Tan, Xin Liu, Bin Xu, Han Zhao, Xiaowei Liu, Jun Shi, Tracy Yu, Hui Gao, Huiji Gao, Liwei He, Michael Kinoti, Stephanie Moyerman, and Sanjeev Katariya Introduction Planning a trip on Airbnb rarely happens in a single session. A guest searching for a place to stay in San Francisco might browse dozens of listings over several days, leaving behind a trail of views. Typically, over a period of years, that same guest will have accumulated many previous bookings, reviews, and the occasional cancellation. Taken together, these events reveal a great deal about what that guest values in a stay. For years, Airbnb's search ranking captured this through hand-crafted features: aggregated statistics such as total past bookings or average listing price. These worked well, but as the feature count grew into the hundreds, the approach became harder to scale and increasingly limited in expressiveness. In this blog post, we describe how we built a sequence modeling system that encodes the full guest journey using a Transformer, learning richer representations of guest preferences to deliver more personalized search results. An example of a guest journey, which is typically long, exploratory, and complex.Challenges Event sequences per guest present three core challenges. First, they are dominated by listing views, which account for the vast majority of all events -- some guests accumulate hundreds of thousands of them -- making raw sequences computationally intractable to model directly. The distribution of event types, with the majority being listing views. Second, unlike social media platforms, which optimize for engagement, Airbnb optimizes for booking conversion. Bookings are rare, compared to events, and deliberate, whereas a listing view could reflect genuine intent or simply idle browsing. Building a model that generalizes ## Building a Transformer-Based Category Recommender at Thumbtack DevFeed: [Building a Transformer-Based Category Recommender at Thumbtack]() Original publisher: [Read original article]() Author: Andrew Morss Published: 2026-07-20T20:23:27Z Content type: article Language: en Sources: [Thumbtack Engineering - Medium]() Topics: [recommendation systems](), [Transformer](), [Machine learning](), [email](), [push-notification]() Tags: [catalog](), [data-science](), [email](), [engineering](), [machine-learning](), [model](), [push-notification](), [recommendations](), [recommender-systems](), [torch]() ### AI overview This engineering article describes Thumbtack's category recommender, which selects home-service categories for push and email notifications. It explains how the system addresses popularity bias through negative sampling and how Thumbtack's small category catalog and sparse user activity shape the recommendation problem. ### Source excerpt A look at compensating for popularity bias in recommender systems using negative sampling strategies By: Andrew Morss, Senior Applied Scientist Introduction A recommender system is a machine learning model that, given a user and a catalog of items, predicts which items that user is most likely to want. Recommenders set your YouTube playlist, determine what items Amazon suggests for you, push you songs on Spotify and customize your Steam store. If you're a homeowner, Thumbtack's recommender systems can suggest home projects for you such as house cleaning or lawn mowing. Thumbtack connects users with professionals spread across almost 500 different job categories. Job categories are a kind of occupational bucket that Thumbtack uses as a taxonomy to categorize local service professionals (pros). Pros have the opportunity to choose categories of jobs that they want to receive leads on; examples include things like "Electrical and Wiring Repair", "Local Moving" or "Snow Plowing". Thumbtack's Lifecycle marketing team uses a category recommender to choose which categories to surface to users in our push and email notifications. Choosing the right ones is critical. A well-chosen suggestion is how a customer starts a search for a new project, finds the right service professional, and ultimately gets work done. An example of a push notification and an in-email category recommendation. The Problem Space: Category Recommendation There are several features of category recommendations at Thumbtack that are fairly unique in the item recommender space. The number of categories that we have is relatively small (~500) compared to the number of items that are usually seen in recommender systems, which simplifies the problem considerably. Usually recommenders require a two stage approach where an initial, low compute, filtering stage reduces the number of items to some "candidate" list of items before a more complicated model ranks that smaller set of items. With a relatively small cat ## Exploring Hierarchical Interest Representation For Meta Ads Deep Funnel Optimization DevFeed: [Exploring Hierarchical Interest Representation For Meta Ads Deep Funnel Optimization]() Original publisher: [Read original article]() Author: Yuhui Ouyang; Di Wang; Sreedal Menon; Jie Tian Published: 2026-07-15T17:00:52Z Content type: article Language: en Sources: [Engineering at Meta](), [Meta AI Research]() Topics: [Optimization](), [Embeddings](), [Graphs](), [recommendation systems](), [LLMs](), [Artificial Intelligence](), [Transformer](), [data]() Tags: [ads](), [ai](), [ai-research](), [data](), [embeddings](), [engineering](), [frontier-ai](), [generative](), [learning](), [meta](), [optimization](), [personalization](), [reasoning](), [research](), [retrieval](), [scale]() ### AI overview Meta describes Hierarchical Interest Representation, an upstream system that learns unified embeddings for users, advertisers, products, and services. It combines graph learning, multimodal content processed through LLMs, engagement signals, and self-supervised distillation to improve personalization, retrieval, ranking, and deep-funnel advertising optimization. ### Source excerpt Hierarchical Interest Representation is a research area for Meta Ads. We're exploring an upstream representation layer over the universe of Ads entities - users, advertisers, products, services - learning unified embeddings that connect users' inferred interests with the breadth of what advertisers offer in their deep funnel ads. The innovations in Hierarchical Interest Representation are [...] Read More... The post Exploring Hierarchical Interest Representation For Meta Ads Deep Funnel Optimization appeared first on Engineering at Meta. ## How to Survive the Recommendation Engine Question DevFeed: [How to Survive the Recommendation Engine Question]() Original publisher: [Read original article]() Author: Raul Junco Published: 2026-07-11T11:35:16Z Content type: tutorial Language: en Sources: [System Design Classroom]() Topics: [recommendation systems](), [Architecture & Design]() Tags: [architecture](), [how-to](), [recommendations]() ### AI overview A tutorial on answering recommendation-engine system design questions by focusing on the surrounding system rather than only the matching algorithm. It emphasizes separating recommendation generation from recommendation serving and addressing behavior data, precomputation, latency, stale results, missing data, cold starts, and fallbacks. ### Source excerpt Most engineers focus on the matching algorithm. The real test is separating generation from serving. ## Worth Reading 070126 DevFeed: [Worth Reading 070126]() Original publisher: [Read original article]() Author: Russ Published: 2026-07-01T12:59:45Z Content type: article Language: en Sources: [rule 11 reader]() Topics: [recommendation systems](), [Cybersecurity](), [systems](), [Rocket]() Tags: [analysis](), [article](), [china](), [cybersecurity](), [space](), [spacex](), [systems](), [worth-reading]() ### AI overview A roundup of developer and technology reading covering recommender-system task analysis, SpaceX and orbital debris, semiconductor cooling reproducibility, and the evolution of cybersecurity risks and defenses. ### Source excerpt A better, more nuanced understanding of tasks in recommender systems can help minimize user costs across the entire recommendation process. The current estimate of the world's population is 8.264 billion people, so the share price of SpaceX is currently at a phenomenal USD $261 per head. The analysis from space monitoring firm LeoLabs, provided to Breaking Defense, found that from January 2021 to January 2025 China has abandoned 51 spent rocket bodies in LEO above 650 kilometers (about 404 miles) in altitude, more than doubling the number for the previous five years to bring the total to 96. A big problem, Pauzauskie said, revolves around reproducibility. So far, labs haven't been able to show that they can consistently cool semiconductors. This article recounts the evolution of modern computing systems to provide an analysis of security risks and their evolution, with a past and present look at key cyber-defense innovations as well as a perspective on future cybersecurity hard problems. ## GenPage: Towards End-to-End Generative Homepage Construction at Netflix DevFeed: [GenPage: Towards End-to-End Generative Homepage Construction at Netflix]() Original publisher: [Read original article]() Author: Netflix Technology Blog Published: 2026-06-29T13:01:02Z Content type: article Language: en Sources: [Netflix](), [Netflix TechBlog - Medium]() Topics: [Netflix](), [recommendation systems](), [Large Language Model](), [Machine learning](), [Reinforcement learning](), [Transformer]() Tags: [ai](), [app](), [diversity](), [engineering](), [feature-engineering](), [generation](), [generative](), [large-language-models](), [machine-learning](), [model](), [netflix](), [optimization](), [ranking](), [recommendation-system](), [reinforcement-learning](), [rl]() ### AI overview This Netflix developer article introduces GenPage, a generative approach that uses a single autoregressive model to construct a personalized homepage by generating recommendation rows, entities, and layout together. It describes replacing a multi-stage recommendation pipeline with end-to-end modeling and using reinforcement learning to optimize whole-page rewards, including interactions such as diversity and the balance between rows. ### Source excerpt Authors: Lequn Wang, Jiangwei Pan, and Linas Baltrunas Figure 1. Autoregressive homepage generation. GenPage builds a Netflix homepage one row or entity at a time, each one conditioned on what's already on the page and the user's context.Introduction The Netflix homepage is the first thing users see when they open the app and the primary way they discover content to enjoy. Almost every part of it is personalized, including which rows appear, which entities show up within those rows, and how everything is arranged on the page. Constructing that homepage is a genuinely hard problem. It is not simply producing one ranked list. The homepage is a structured, two-dimensional layout, made up of recommendation rows and the entities within them. Here, an entity can be a movie, show, game, live event, or other recommendable item. Each choice can affect the value of the others. Traditionally, it is built through a complex, multi-stage pipeline, with separate components for candidate generation and ranking at both the row and entity levels. We saw an opportunity to rethink this design. Large language models have shown that a single generative model can perform diverse tasks just by generating a response to a prompt. Inspired by this prompt-response paradigm, we trained a single generative model to build the homepage by directly answering one question: Given everything we know about this user and this request, what homepage should we generate to maximize user satisfaction? We call this approach GenPage. It treats the user history and request context as the prompt, and autoregressively generates the entire homepage as the response (Figure 1). Unlike most generative recommenders, such as TIGER, HSTU, and OneRec, which generate flat ranked lists, GenPage generates the rows, entities, and layout together. This shift is motivated by several goals: End-to-end modeling. A single transformer model that constructs the page from raw input signals can replace a complex multi-stage recommen ## Distilling Long-Tail User Behavior into Scalable Embeddings for Job Search DevFeed: [Distilling Long-Tail User Behavior into Scalable Embeddings for Job Search]() Original publisher: [Read original article]() Author: Marsan Ma Published: 2026-06-03T23:28:10Z Content type: article Language: en Sources: [Indeed]() Topics: [recommendation systems](), [Embeddings](), [Deep learning](), [Latency](), [Job]() Tags: [big-data](), [data-science](), [deep-learning](), [embeddings](), [engineering](), [feature-store](), [job-search](), [latency](), [machine-learning](), [performance](), [ranking]() ### AI overview Indeed describes a user behavior modeling system for job search that learns from long-term user histories offline, distills them into fixed-length embeddings, and serves them through a feature store for use by online ranking and recommendation models. The approach is designed to preserve rich behavioral signals while meeting latency and cost constraints. ### Source excerpt Authors : Marsan Ma, Nikhil Lopes, Raj Amrit, Hong Lu, Dipankar Biswas, Trent KyonoLeadership: Iris Wang, Madhu Kurup Recommendation and ranking systems power many of the most important experiences on large internet platforms. Yet the models that run in production are rarely the largest models we can train. They are usually compact, latency-sensitive supervised models [...] ## Semantic IDs: Product Understanding at Scale DevFeed: [Semantic IDs: Product Understanding at Scale]() Original publisher: [Read original article]() Author: Shrikar Archak Published: 2026-06-02T16:58:00Z Content type: article Language: en Sources: [Instacart]() Topics: [recommendation systems](), [data](), [systems]() Tags: [embedding](), [ids](), [machine-learning](), [model](), [product](), [recommendation-system](), [recommender-systems](), [retrieval](), [scale](), [vector]() ### AI overview Instacart describes how it built semantic IDs to capture relationships among grocery products that a hierarchical taxonomy misses. The approach uses product embeddings, residual vector quantization, catalog structure, contrastive training, and separate precision and discovery strategies to address cold-start products, sparse categories, and catalog-quality issues. ### Source excerpt Key Contributors: Shrikar Archak, Karuna Ahuja, Soroush Sobhkhiz, Marko Avdalovic, Xiyu Wang, JiChao Zhang, Hao Yan, Chris Hartley Introduction Operating a grocery catalog at Instacart's scale means managing millions of products across thousands of categories. Every product is assigned to a category in our hierarchical taxonomy like "Dairy > Cheese > Parmesan". These categories provide broad classification, but they miss the connections that drive how customers actually shop. For example, a customer is building a cheese board. They've added Parmigiano Reggiano, and now they need accompaniments. Our taxonomy puts it in "Dairy > Cheese > Parmesan," so a category-based system can suggest other parmesan cheeses. But it can't connect them to the Castelvetrano olives in Pantry > Condiments > Olives, the olive tapenade in Deli > Olives Dips and Spreads, or the crudité and pre-assembled cheese tray in Deli > Prepared Meals > Party Trays. These products live in completely different branches of the catalog, with no shared ancestor below "Food." But any customer would tell you they belong together. This cross-category blindness shows up in three ways. Cold start: new products arrive with zero purchase history. We can assign them to the right category, but a category alone can't connect them to the products customers would actually consider alongside them, so they stay invisible. Tail category coverage: recommendation models learn from volume, so they skew toward popular grocery staples. Products in sparse categories lack the interaction data to surface, and the taxonomy gives the model no bridge to related items in other branches. Catalog quality at scale: with millions of products, mislabeling is inevitable -- a protein bar filed under "Candy," a sparkling water under "Soda." A rigid tree has no way to flag these because the only signal is the label itself. In this post, we walk through how we built semantic IDs at Instacart to address these problems: the embedding choices, th ## SilverTorch: Index as Model -- A New Retrieval Paradigm for Recommendation Systems DevFeed: [SilverTorch: Index as Model -- A New Retrieval Paradigm for Recommendation Systems]() Original publisher: [Read original article]() Author: Lei Chen; Yiyi Pan; Ivy Sun; Sha Meng; Cornelia Carapcea; Shilin Ding; Ram Ramanathan; Nipun Mathur; Hong Yan; Lars Backstrom Published: 2026-05-26T16:00:01Z Content type: article Language: en Sources: [Meta ML Applications]() Topics: [recommendation systems](), [Neural Network](), [Architecture & Design](), [systems]() Tags: [compute](), [efficiency](), [engineering](), [meta](), [ml-applications](), [model](), [recommendation-systems](), [research](), [retrieval](), [systems]() ### AI overview Meta introduces SilverTorch, a unified model-based retrieval system for recommendation systems. It replaces separate retrieval microservices with modules in a single neural network, allowing more complex modeling and more candidates to be evaluated within a sub-100-millisecond target. The article reports up to 23.7x higher throughput and 20.9x greater compute cost efficiency than specified baseline approaches. ### Source excerpt We're introducing SilverTorch, a reimagining of recommendation systems that unifies all retrieval components for user generated content under a unified architecture. SilverTorch shows up to 23.7x higher throughput compared to the state-of-the-art approaches. It's also showing 20.9x more compute cost efficiency compared to a CPU-based solution while also improving accuracy. Our research paper, "SilverTorch: A [...] Read More... The post SilverTorch: Index as Model -- A New Retrieval Paradigm for Recommendation Systems appeared first on Engineering at Meta. ## Making User-Sequence Data More Cost-Efficient, Faster, and Easier to Use DevFeed: [Making User-Sequence Data More Cost-Efficient, Faster, and Easier to Use]() Original publisher: [Read original article]() Author: Pinterest Engineering Published: 2026-05-21T16:01:00Z Content type: article Language: en Sources: [Pinterest Engineering Blog - Medium]() Topics: [data](), [recommendation systems](), [datasets](), [Machine learning](), [Embeddings](), [Feature Engineering](), [Inference](), [Latency](), [systems]() Tags: [data](), [data-infrastructure](), [datasets](), [embeddings](), [engineering](), [feature-engineering](), [inference](), [latency](), [machine-learning](), [pinterest](), [production](), [recommendation-system](), [recommendation-systems](), [retrieval](), [systems](), [train](), [transformers]() ### AI overview Pinterest describes a redesign of its user-sequence platform for ranking, retrieval, and recommendation workloads. The article explains how enriched event sequences support training datasets, offline analysis, online inference, and latency-sensitive production use cases, with goals of reducing cost, improving extensibility, and simplifying debugging. ### Source excerpt Authors (listed alphabetically) Ads Feature Engineering Infra team: Ajay Venkatakrishnan, Le Zhang Core ML Infra team: Eric Shang, Pihui Wei ML Data team: Connor Votroubek, Yi He User Understanding team: Camilo Munoz, Simin Li If you work on ranking, retrieval, or recommendation systems, you've probably asked for some version of the same thing: "Give me the last N meaningful actions this user took, with the right enrichments, in a format that's easy to train and serve ML models." On paper, that sounds simple. In practice, "user sequences" often become one of the most expensive and fragile parts of the ML data stack. They end up powering everything from training datasets to offline analysis and online inference, so they need to be fresh and complete at the same time. They must remain consistent as you add new events and enrichments. And they have to do all of this while serving latency-sensitive production workloads. This article walks through how we redesigned our user-sequence platform to make these sequences cheaper to run, faster to extend, and easier to debug, while still supporting demanding production use cases. What We Mean by "User Sequence" In this context, a user sequence is an ordered list of recent, relevant events for a user, along with the enrichments (signals) attached to each event. Here, enrichments mean all the extra signals we attach to raw events, so they're useful for models: embeddings (for example, Pin or query representations), contextual features (such as surface, device, or country), and derived attributes or counters that describe how the user interacted with a piece of content over time. A concrete example helps. Imagine a sequence made up of the last 500 engagements a user had with Pinterest Pins. Each event in that sequence might carry a timestamp, an action type, the surface where the action occurred, and a handful of embedding features or categorical attributes. As a data primitive, user sequences are powerful. They capture temporal b ## How Whatnot's Hourly ML Feature Pipeline Safely Handled a Failed Update DevFeed: [How Whatnot's Hourly ML Feature Pipeline Safely Handled a Failed Update]() Original publisher: [Read original article]() Author: Whatnot Engineering Published: 2026-05-07T16:01:01Z Content type: article Language: en Sources: [Whatnot Engineering]() Topics: [Machine learning](), [recommendation systems](), [datasets](), [data](), [real-time](), [Processes]() Tags: [ai](), [data](), [data-engineering](), [feature](), [feature-engineering](), [feature-store](), [machine-learning](), [ml](), [pipeline](), [ranking](), [real-time](), [recommendation-system](), [recommendations](), [run]() ### AI overview Whatnot describes how its hourly ML feature pipeline handled a bad pull request that caused a feature group to produce zero rows. The online feature store served last-known-good features for up to two days, anomaly monitors detected the row-count drop, and the team reverted the change before the freshness gap became significant. The post shares lessons about safe degradation and operating frequently updated recommendation features. ### Source excerpt David Kawashima, Miguel Fernandez-Montes Cuberta, Ford Bohrmann | Discovery A bad PR went out just as our hourly feature pipeline was about to run. The pipeline was responsible for recomputing fresh user behavior signals used in livestream ranking on the home feed. These features capture recent user intent: watches, bids, orders aggregated across users, sellers, and categories. A missed feature update means that the system stops responding to user signals immediately. This PR release caused one of our feature groups to suddenly produce zero rows ... but nothing broke! Even though a critical feature group for the model stopped updating, the online feature store continued serving the last-known-good features, buffered by a 2-day TTL and nothing broke explicitly. Our anomaly monitors went off immediately, flagging the drop in row counts. The team identified the bad change and reverted it before the freshness gap grew large enough to matter. No corrupted data reached production. The system absorbed the failure and degraded safely. This graceful degradation didn't happen by accident; it was the result of hardened processes from years of lived experience, seeing the way these systems can degrade in obvious or subtle ways. In this post, we share some of those lessons learned. The system: hourly feature pipeline Whatnot is a highly dynamic marketplace where buyers' preferences change at a moment's notice. A buyer might research a laptop, join a football-card break a few hours later, and buy chocolate before logging off for the night. Daily feature updates capture long-term preferences but miss intra-day shifts. To keep up with these rapidly changing buyer interests (in 2025) we moved from daily to hourly updates for features that power our ML feed ranking models. Fresher features are a critical component for capturing the live nature of the Whatnot marketplace, where the inventory available to purchase changes in real time. Moving from daily to hourly updates was a meaningful ## Empowering Carrot Ads with Domain Adaptive Learning DevFeed: [Empowering Carrot Ads with Domain Adaptive Learning]() Original publisher: [Read original article]() Author: Xiyu Wang Published: 2026-05-04T19:11:17Z Content type: article Language: en Sources: [Instacart]() Topics: [recommendation systems](), [Artificial Intelligence](), [data](), [Optimization](), [Website]() Tags: [advertising](), [ai](), [data](), [domain](), [machine-learning](), [optimization](), [ranking](), [retail]() ### AI overview The article explains how Instacart applies Domain Adaptive Learning to Carrot Ads to address the cold-start problem for new retail partner websites. It describes transferring knowledge from Instacart Marketplace data to target domains to improve click-through-rate prediction and ad ranking with limited partner-specific interaction data. ### Source excerpt Authors: Trey Zhong, Xiyu Wang Contributors: Joseph Haraldson, Sharad Gupta, Sarah Lamacchia Introduction Carrot Ads is Instacart's omnichannel retail media solution that allows retailer partners to build and scale their own advertising businesses on either their owned-and-operated (O&O) websites and apps or their whitelabel Storefront hosted by Instacart. Carrot Ads empowers retailers and CPG brands to accelerate revenue, while improving the customer experience, engagement and Ads return on investment. It features enterprise-grade infrastructure, AI-powered optimization, years of proprietary first-party data and flexibility to choose from retailer-sourced Ads demand, Instacart-sourced demand from 7,500+ CPG brands, or both. However, onboarding a new partner onto Carrot Ads introduces a key challenge: the 'cold start' problem, where limited historical interactions make it difficult to predict user behavior accurately. To serve performant ads, our systems rely on predicting a user's Click-Through Rate (CTR) to generate a ranking score. On the Instacart Marketplace, we have billions of historical signals to train a model to do so. But when a partner launches a new ads experience on their O&O e-commerce site, there is often little to no interaction history for that property, so training an accurate model becomes challenging. User behavior can vary dramatically between websites -- for example, browsing patterns on a grocery site differ from those on a pet supply or electronics site. Training a model from scratch for a new domain is data hungry. Conversely, directly deploying Instacart's existing Marketplace model often fails to capture the nuances of the partner's specific inventory and user base. To address this, we developed a Domain Adaptive Learning approach that transfers knowledge from Instacart's data-rich environment to new partner environments. By treating the Instacart Marketplace as a source domain and the partner's website as a target domain, we can transfer [Next page]()