Making recommendations explainable
summary
Replaces opaque user embeddings with a list that stores positive publication IDs and builds a lookup mapping each positive to its candidate set; filters pairs with crowd-sourced labels, trains a relevance model, and ranks candidates by an attractiveness model that uses conditional CTR plus meta features. Prunes growing positive histories with a determinantal point process and allocates candidate quotas with a bandit-style proportional split. Reports higher feed relevance and much better developer traceability for debugging.
AI-written summary · about
chapters (15)
- 0:00Welcome & What You’ll Learn: Making Recommendations Explainable
- 1:03What “Explainable Recommendations” Mean (User View)
- 1:29Why Change the System: User Reports & Developer Traceability
- 2:03The Classic Recommender Pipeline Overview
- 2:46Embeddings 101: Content/Collaborative Filtering Basics
- 3:49User Profile Stage: Why Single Embeddings Are Hard to Explain
- 5:19Candidate Selection with KNN/HNSW: Speed vs. Quality Issues
- 7:03Ranking + Business Logic: Scoring and Diversifying the Feed
- 8:08New Explainable Approach: Store Positive Publication IDs
- 9:09Building the Explainability Index: Relevancy via Labels & Models
- 10:25Ordering Candidates: Attractiveness (CTR → Predictive Model)
- 11:49Generating the Explanation Text Users See
- 12:25Keeping Positives Manageable: Diversity + Ranking Positives
- 14:28Smarter Candidate Quotas with Bandit-Style Allocation
- 15:13Wrap-Up: Transparency, Better Relevance, Better Debugging
related talks
25mBuilding a recommendation system for EPAM systems
23mML-Powered Search & Recommendation 101
12mExplaining recommendations in an interactive hybrid social recommender
39mMachine Learning for Relevance and Serendipity
41mRecommender Systems: Crafting Effective Personalised Experiences
38mEmbeddings! Embeddings everywhere!
01h 21mUnlocking Personalization: A Deep Dive into Modern Recommendation Algorithms
57mHands-On with LLM-Powered Recommenders: Hybrid Architectures