Loading…
Topic
Recommendation Systems
54 posts about Recommendation Systems. Every summary links to the original.
- 54 posts
- 21 companies
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Newest first
Posts
Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI
Meta introduced GEM, the industry’s largest recommendation system foundation model, and its 2026 generation shifts from isolated per-domain optimizations on Facebook, Instagram, and Reels to a single central foundation model. The Kunlun architecture, co-designed with the latest GPU hardware, enables 2.5x model complexity and demonstrates LLM-like scaling laws with block-aligned sparse attention reducing sequence complexity. The LoopFM post-training technique transfers GEM’s internal representations to downstream models, doubling intelligence transfer at no additional compute cost or latency. GEM delivered 18.9% conversion improvement and 12% ad click-through rate improvement across all Meta ad surfaces and funnel stages, including 5.8% conversion uplift on Facebook, 4.1% on Instagram, and 1.6% on off-site conversion. Training efficiency doubled to 20-25% model FLOPS utilization while scaling total training FLOPs by 4x across thousands of GPUs using custom kernels and 5D parallelism.
Is machine learning hard? Not always - The Intercom Blog
The post argues that machine learning does not always require advanced mathematics and, as described by Intercom’s Fergal Reid, is about systems that improve with more training data. It contrasts a naive keyword-matching approach in Ruby for a ‘You might also like’ feature, which produced irrelevant results, with a basic machine learning method using K-Means clustering. The machine learning approach first represents each blog post as a binary vector indicating the presence of each word from the entire corpus. The K-Means algorithm then iteratively assigns each post to the nearest cluster center and recalculates centers as the average of assigned posts until assignments stabilize. The post demonstrates that roughly 40 lines of Ruby code can produce relevant recommendations and encourages readers to try libraries like Scikit Learn.
Turning Opportunistic AI Inference Capacity Into Better Ads Recommendations
Meta built a system that aggregates fragmented idle inference capacity from a globally constrained fleet into a shared opportunistic pool to enhance ads recommendations. Idle capacity arises from four mismatches: peak provisioning with demand variation, model development timing with capacity delivery, regional resource ratios, and accelerator types. Two elasticity levers exploit this preemptible, volatile capacity: Rank More Ads increases candidates evaluated by ranking stages, using both CPU and accelerator surplus, and Serve Better Models routes more requests to a more powerful model, using accelerator-only surplus. The system protects baseline serving through isolated routing, minute-scale feedback loops that adjust candidate volume and model coverage, and millisecond-scale value-aware load shedding that removes lower-value work during sudden preemption. This allows optional work to expand and contract with opportunistic capacity, driving better ad relevance and advertiser outcomes without requiring additional capacity orders.
Real-Time Retail Intelligence: Building E-Commerce Recommendations with Lakebase and AI Search on Databricks
The post presents a reference architecture for building a production-grade e-commerce recommendation system on Databricks, based on a real-world implementation for a leading fashion e-commerce platform in Asia serving over 1 million monthly active users across 100,000+ SKUs. The system ingests approximately 1,000 events per second via Lakeflow Connect’s Zerobus into Unity Catalog Delta tables within a medallion architecture. Serving uses two paths: Path A for pre-computed batch recommendations via key-value lookup from Lakebase online tables, and Path B for real-time session-aware scoring using AI Search and a LightGBM model in a 3-stage funnel. The architecture unifies ingestion, feature management, model training, vector retrieval, and serving on a single governed platform, with cold-start handling for new users and products. Expected outcomes include low 2-digit millisecond responses, daily-fresh recommendations, and unified governance via Unity Catalog.
MAPS: Netflix’s Multimodal Asset Personalization at Scale
Netflix treats artwork and video previews as personalized assets, but ID-based models struggle with newly launched titles and assets that have little interaction history. The systems encode artwork with 768-dimensional CLIP image embeddings and combine those vectors with learned asset IDs, allowing visual preferences to transfer across titles and enabling one model to serve five canvases. Reward-based weighting balances pooled canvas data by the long-term value of interactions, while inverse propensity scoring on randomized exploration traffic supports offline evaluation. For video previews, MediaFM combines visual, audio, and text signals and ranked ahead of SeqCLIP in a linear probe, offline IPS evaluation, and online A/B testing; the Netflix Embedding Store supplies consistent embeddings to production models.
Scaling Conditional Learned Retrieval for Pinterest Home Feed
Pinterest home feed candidate generation relies on large-scale retrieval, but standard two-tower models produce single embeddings that struggle to capture diverse concurrent user intentions. To solve this, Pinterest expanded Conditional Learned Retrieval (CLR) to generate condition-aware embeddings conditioned on explicit interest, Pin, and Board contexts. The team evolved model representations by appending condition tokens to user sequences in transformers, integrating the PinFM foundation model, and introducing hierarchical semantic IDs to improve cold-start coverage. On the serving side, engineers consolidated duplicate feature requests and shifted inference to GPU instances using NVEmbed under a one-query-to-many-documents paradigm. These infrastructure and modeling changes unlocked new condition types, achieved seven-figure cost savings, and reduced p90 model retrieval latency from 80ms down to 12ms.
Inside LinkedIn's cognitive memory agent for agentic personalization
LinkedIn built a cognitive memory agent to deliver personalized experiences across its recruiter workflows and hiring assistants. Interactions between recruiters and agents often reveal recurring preferences, such as candidate criteria, role adjustments, and task habits. To preserve and manage this state throughout the entire lifecycle, the system organizes user data into a multi-tiered memory architecture. This structure comprises conversational, semantic, episodic, and procedural memory tiers that track everything from ongoing sessions to cross-session behaviors and workflow methods. LinkedIn continues to research memory abstractions, dynamic session compaction, virtual file system storage layers, and robust evaluation approaches for attribution across complex multi-source application agents.
From Prediction to Action: How to Turn AI Outputs Into Decisions
Salesforce addressed an operational challenge where sellers faced roughly 12,000 dashboards and over 20 applications outputting machine learning predictions without clear next steps. The engineering team reframed machine learning outputs as raw signals rather than standalone answers. To bridge the gap between assessment and action, they built a Next Best Action layer that combines model signals, business logic, and contextual institutional knowledge into actionable recommendations. They integrated this layer with an AI agent using Model Context Protocol (MCP) tool contracts, enabling dynamic discovery and explicit handling of missing data. Finally, rather than introducing a separate dashboard destination, the agent serves on-demand recommendations directly inside Slack where sellers already collaborate.
Connecting retail demand planning to campaign and store execution
Retail and consumer goods teams often split performance review, demand planning, marketing activation, store execution, and measurement across disconnected systems, making it difficult to connect analysis with results. The Databricks demo presents a unified application built on data unification, governance, and intelligence, combining point-of-sale, loyalty, supply chain, media spend, inventory, and store-operations signals with role-specific access. Users can move from KPI and forecast-versus-actual analysis to SKU or store what-if planning, audience selection, campaign activation, replenishment, price-tag updates, assigned tasks, and post-campaign measurement. Genie supports natural-language querying, while specialized agents and a coordinating supervisor connect sales insights, demand planning, audience building, in-store operations, and measurement. The stated design goal is a governed, shared workflow that lets teams test decisions, execute them across stores, and evaluate outcomes without losing operational context.
How Generative Recommenders Are Redefining RecSys at Scale
Generative recommenders (GRs) recast recommendation as next-item or next-action sequence prediction conditioned on user histories, replacing traditional embedding-similarity objectives and potentially unifying retrieval and ranking. The post presents HSTU, which uses timestamp-ordered interleaved items and actions, SiLU-based attention weighting, relative attention bias, and gating, alongside Semantic IDs, which derive a smaller vocabulary through hierarchical clustering of item embeddings. For production, recsys-examples combines TorchRec, DynamicEmb, Megatron-Core, recommender-specific KV and decoding components, while nv-embedding-cache provides tiered GPU, CPU, and remote embedding storage. On an H100, the Semantic ID path was reported as 1.85x higher in online throughput and 46% lower median latency than SGLang under the stated long-context, beam-search workload, and DLRM v3 reached 99,997 queries/sec.
From User Sequences to Scaling Laws: A Multi-Stage Architecture for Meta’s Ads Ranking
Meta’s ads ranking systems process millions of candidates per second under millisecond latency budgets, creating tension between richer sequence models and efficient serving. The post presents a multi-stage architecture that moves long-history user modeling into asynchronous offline transformers, caches user-level embeddings, and combines them with fresh user and ad signals in an online ranking model. Dense tokenization places sparse features and behavioral sequences in one vocabulary, while target-aware multi-head attention learns higher-order interactions between historical behavior and the candidate ad. On real-world ads traffic, performance follows a log-linear relationship with compute across depth, enrichment, width, and sequence length, with balanced growth and diverse sequences identified as scaling levers. Together with broader model innovations, the approach contributed to a cumulative 6% conversion lift on Instagram, 3% on Facebook, and 3.5% lift in Facebook ad clicks, and serves as a core component of GEM.
GEM Training: How Meta Doubled the Efficiency of Its LLM-Scale Ads Foundation Model
Meta’s Generative Ads Recommendation Model (GEM), a hybrid foundation model for ads recommendations on Instagram and Facebook, trains across several thousand latest-generation GPUs and combines trillions of sparse embedding parameters with billions of dense parameters. The post frames its engineering problem around recommendation-specific jagged and asymmetric sequences, memory-bound operations, numerical sensitivity, communication overhead, activation pressure, and load skew that make conventional LLM training techniques insufficient. Its co-design approach combines custom kernels such as Jagged Flash Attention (JFA), BlockAttention, and Generalized Dot-Product Attention (GDPA), mixed ultra-low-precision training including MXFP8 attention and MLP, topology-aware 5D parallelism, SM-free collectives, and memory and load-balancing techniques. The reported result is a doubling of end-to-end training efficiency to 20–25% Model FLOPs Utilization (MFU) while total training FLOPs increased fourfold over 12 months; Base Batch Shuffling (BBS) alone delivered a reported 4% efficiency gain, with 4% QPS improvement and 4% peak memory reduction.
GenRec: Towards LLM-Native Recommendation at Netflix
Netflix developed GenRec, an LLM-backed recommendation ranker built by adapting an internal foundation model for large-scale personalization. Traditional recommendation stacks rely on thousands of hand-crafted features and specialized architectures that are costly to maintain across diverse content types. To replace manual feature pipelines, GenRec verbalizes user histories, metadata, and contexts into natural-language prompts and trains with multi-objective losses, including catalog-aware ranking and reward-weighted alignment. At inference time, the model executes in prefill-only mode on vLLM without decoding text. In large-scale online A/B testing against a mature production ranker, GenRec achieved statistically significant improvements in short-term and long-term metrics while using fewer labeled examples.
Pinner Progression: Better Use-Case Representation Driving Weekly Active User Growth at Pinterest
Pinterest designed the Pinner Progression initiative to shift its home feed recommendation pipeline from short-term engagement optimization toward long-term user retention. Standard ranking systems model user actions as immediate sequential features, failing to capture the evolving lifecycles of distinct user habits. To address this gap, the platform introduced User Interest Clusters (UICs), which cluster a user's engaged Pins in OmniSage embedding space using dynamic cluster counts and stateful lifecycle metadata. Pinterest integrated a UIC-aware penalty directly into Sliding Spectrum Diversity (SSD) scoring by matching Pin embeddings to cluster medoids via cosine similarity. Online experiments demonstrated that balancing use-case representation increased interaction diversity, extended session lengths, and delivered meaningful engagement gains across the platform.
Personalizing Airbnb search by learning from the guest journey
Airbnb replaced hundreds of hand-crafted ranking features with a Transformer-based sequence model that captures both long-term booking history and short-term browsing behavior. To keep latency low and throughput high, sequence representations are generated via daily batch jobs and combined with real-time queries using a co-trained setwise ranker.
Building a Transformer-Based Category Recommender at Thumbtack
Thumbtack developed a transformer-based category recommender to personalize email and push notification suggestions across roughly 500 home service categories. Because user activity is sparse while contextual signals like location and climate are rich, the architecture represents search history and context features as sequences of tokens evaluated against candidate category tokens using cross-attention. Initial training with naive cross-entropy suffered from severe popularity bias, skewing recommendations toward head categories and yielding a low tail hit rate. To resolve this issue, the team implemented adaptive mixed sampling, combining in-batch and random negative candidates and keeping only top-scoring hard negatives for cross-entropy loss computation. This method raised tail hit rate from 2.1% to 28.4%, improved overall HR@10 to 54.6%, and yielded a statistically significant 5% lift in requests per user in production A/B testing.
Exploring Hierarchical Interest Representation For Meta Ads Deep Funnel Optimization
Hierarchical Interest Representation is an upstream representation layer for Meta Ads that addresses sparse deep-funnel feedback by connecting users’ inferred interests with advertisers, products, services, and other ads entities. It models a typed, heterogeneous engagement graph with a transformer architecture, sparse attention for long-range relationships, bias-aware attention, self-supervised cross-view distillation, and multimodal content features processed through vision or language models. The system projects the raw graph into learned interest primitives and cascading super-graphs, producing universal embeddings and Bag-of-Meaning tokens at multiple granularities. Temporal cutoffs and chronological splits prevent future-information leakage, while a pipelined online graph engine reportedly achieved a 30x wall-clock speedup over a synchronous baseline during testing. Trained end-to-end on billions of interactions, the representations are designed for personalization, retrieval, supervision, and ranking integrations including GEM, Andromeda, and the Adaptive Ranking Model.
GenPage: Towards End-to-End Generative Homepage Construction at Netflix
Netflix traditionally constructs its structured two-dimensional homepage through a complex, multi-stage recommender pipeline that separates candidate generation and ranking across rows and entities. To simplify this architecture and optimize directly for whole-page user satisfaction, Netflix developed GenPage, an end-to-end generative transformer model that autoregressively builds the entire homepage from raw tokenized context. The system relies on a domain-specific tokenizer to compress engagement history, context injection for cold start, hybrid row decoding to minimize decoding steps, and reinforcement learning post-training. In online A/B testing against the production baseline, GenPage delivered statistically significant gains in core user engagement metrics while cutting end-to-end serving latency by 20 percent. Offline evaluations further showed that enriching context prompts improved recommendation quality more effectively than increasing model capacity in the current operating regime.
Achieving Near-Linear Training Scalability for Pinterest’s Foundation Models
Pinterest’s foundation ranking models store approximately 99% of their parameters in embedding tables, causing cross-node NCCL all-to-all communication to severely bottleneck multi-node distributed training. Initial multi-node scaling stalled at 1.21x on four nodes even after adopting AWS Elastic Fabric Adapter for OS-bypass networking. To resolve this, engineers profiled GPU activity with PyTorch Profiler and systematically addressed communication bottlenecks across multiple layers. The team implemented FP8 quantized communications using FBGEMM, balanced table sharding across GPUs, halved embedding dimensions while doubling row counts, and adopted a 2D parallel communication topology alongside PyTorch Distributed Checkpoint. These combined interventions achieved near-linear scaling factors of 3.9x on four nodes and 7.5x on eight nodes across 64 GPUs.
What is Vector Search?
Vector search retrieves results by comparing embeddings that represent meaning across text, images, audio, and other content rather than matching exact words. A model creates embeddings, an index stores them for fast similarity search, and a query embedding is matched against the index using nearest-neighbor methods. Exhaustive k-nearest neighbor search can become too slow at millions of items, so production systems commonly use approximate nearest neighbor search, trading some precision for speed. The guide positions vector search behind semantic search, RAG, recommendations, and multimodal or cross-language retrieval, while hybrid search combines dense and sparse vectors, keyword results, metadata filtering, and reranking to improve reliability. Quality depends on embeddings, filters, index freshness, and infrastructure, with vector search requiring more memory and compute; Databricks AI Search is presented as a managed service supporting these capabilities and Unity Catalog governance.
Related topics
All topics →Weekly digest
The best engineering reads, once a week.