Technology Trends in RecSys

Kirill Khrylchenko25 May 2024

Problem

A team that runs an industrial recommender (feed, marketplace, streaming) sees a flood of papers on LLMs, scaling, graphs and reinforcement learning (RL). It is hard to tell which directions are already proven in production and which are still open. Picking wrongly costs months of engineering on ideas that do not pay off yet.

Example from the article: classical recommenders know nothing about the outside world, do not solve cold start, are black boxes and give users no way to steer the feed. LLMs address exactly these gaps, but they are expensive and unstable.

Solution in brief

The talk maps six trends seen in industry papers: LLMs, neural ranking, scaling, user-history analysis, graphs and RL. For each it gives the core idea, which companies are active, and what remains unsolved. The most mature are neural ranking and history transformers (the TransAct architecture from Pinterest is widely reused). LLMs, scaling and RL are still early.

Questions this case answers
  • How to tell which research directions in ranking are already proven in production and which are still open before spending engineering months on them
  • How to replace a gradient-boosted ranker with a neural model over the same features and long interaction histories
  • How to decide which modern techniques to adopt in an industrial recommender system
  • How to move recommendation ranking from gradient boosting to a neural network

Tools and methods

  • LLM for recommendations — feature extraction, text signals, candidate generation and ranking
  • Neural-network ranking with multi-task heads — replaces gradient boosting
  • Transformers over user history — long, time-aware, multimodal sequences
  • Graph neural networks and link prediction — use the whole user-item-query graph
  • Bandits and RL — exploration, long-term reward, multi-stage systems
Continue reading — it's free

Google or a one-time email code. No password.