Industry examples, exploration strategies, warm-starting, off-policy evaluation, and more.
Introducing randomness and/or learning from inherent randomness to mitigate position bias.
Thinking about recsys as interventional vs. observational, and inverse propensity scoring.
Simple baselines, ideas, tech stacks, and packages to try.
An overview of system design, candidate retrieval, and ranking, with industry examples.
Focusing on long-term rewards, exploration, and frequently updated item.
Why real-time RecSys? What does the system design look like in industry? How to build an MVP?
Breaking it into offline vs. online environments, and candidate retrieval vs. ranking steps.
A whirlwind tour of bandits, embedding+MLP, sequences, graph, and user embeddings.
Why real-time? How have China & US companies built them? How to design & build an MVP?
Emphasis on bias, more sequential models & bandits, robust offline evaluation, and recsys in the wild.
What I learned about measuring diversity, novelty, surprise, and serendipity from 10+ papers.
Comparing baselines (matrix factorization) against novel approaches using graphs & NLP.
Beating the baseline using Graph & NLP techniques on PyTorch, AUC improvement of ~21% (Part 2 of 2).
Building a baseline recsys based on data scraped off Amazon. Warning - Lots of charts! (Part 1 of 2).