UVENTS

You're voting for

CP4285

Modern Recommendation Systems

Projects

CP4285-01

Prelude: Cold-Start Music Discovery

Music recommenders often favor established songs because those songs have more listening data. Prelude studies how to recommend new and niche songs when their interaction histories are sparse or absent. Using Music4All-Onion, we compare collaborative filtering with models that use song content. A two-tower model performs best overall on warm songs and achieves the highest hit rate on a simulated cold-start test. Item-kNN provides the broadest exposure across cold songs.

CP4285-02

Clew: Session-Based Product Discovery for Anonymous Shoppers

Online stores often know little about shoppers who browse without signing in. We use actions from a shopper’s current visit to recommend products that match their interests. Using real-world shopping logs, we compare methods for selecting and ranking products, then measure how adding variety affects ranking accuracy.

CP4285-03

ForgetMeNot: Poisoning and Recovery in Continually Learned Recommendation Systems

Online shops retrain their recommenders on fresh shopper activity, and sellers can fake that activity to get seen. We asked: once the fake reviews stop, does real data wash them out?

CP4285-04

War of the Embeddings

We recommend news articles based on what a reader has clicked before and what each article says. We ask which signal personalizes news best, for which readers, and at what cost beyond accuracy.

CP4285-05

QCSAR (Query-Conditioned, Session-Aware Ranking): Bringing the "For You" Page into Search

Short-video platforms excel at recommending content but often struggle at search. A user searches for a video, the results miss it, and the same video later appears on their "For You" feed. We believe this happens because platforms reserve heavy, personalised modelling for the feed, while search relies on fast matching between the query and sparse video text. QCSAR asks whether search can borrow the feed's strengths. We build a search pipeline for short videos that combines query relevance with recommendation signals (the user's long-term preferences and their current viewing session) and evaluate it on the KuaiRec dataset. Our goal: search results that answer the query and fit the person asking.

CP4285-06

Squeeze the Balloon: When Debiasing Shifts Bias Elsewhere

Debiasing methods for recommender systems usually fix one bias at a time and are judged on accuracy alone. On the KuaiRand short-video dataset, we test three such methods against all three biases (popularity, exposure and video duration), using randomly shown videos as an unbiased test set. Like squeezing a balloon, removing one bias often pushes another up, and accuracy metrics barely notice.

CP4285-07

A Ranker in Plato’s Cave

A recommender can only rank what it sees—but what if its view of the world is distorted? Through the lens of Plato's Cave, we expose the hidden shortcuts that cause better retrieval and fresher data to backfire, revealing why smarter recommendations require more than just better models.

Comments (Optional)

Token