Dealing with pain points of recommenders in the real world
summary
Alternating least squares shows practical gaps: it handles explicit or implicit data but not both, is sensitive to anomalies, and ignores temporal drift. The implemented solution converts user histories to min hash signatures built from time-dependent bit vectors, uses locality-sensitive hashing to reduce comparisons (example: 270M users down to candidate sets like 100k), and composes device- and period-specific signatures to detect changing tastes, filter transient spikes, and recommend nostalgic-but-not-painful items in a scalable, incremental way.
AI-written summary · about
chapters (9)
- 0:00Start
- 0:09Sophie Watson
- 10:59Anomalous Recordings
- 14:31The Jacquard Index
- 19:13Min Hash
- 22:15Min Hash Signature
- 24:42Locality-Sensitive Min Hash
- 30:01Users Changing Their Opinion
- 33:59Cold Start Problem
related talks
41mRecommender Systems: Crafting Effective Personalised Experiences
58mRecommender Systems – A Walkthrough
52mReinforcement Learning in Recommender Systems: Some Challenges
37mA primer on recommendation systems.
20mLucky Talk: Low-regret Recommendations and Item-item Collaborative Filtering
33mPeople who liked this talk also liked ... Building Recommendation...
16mMaking recommendations explainable
24mWhen Recommendations Systems Go Bad