Phantom Correlation: User Activity Alone Fabricates Item Similarity in Recommender Systems
_reachsumit · x · 2026-10-28
A new paper shows a substantial fraction of item-item similarity in recommender systems is an artifact of user activity variance, not shared taste. Even with fully independent consumption where users differ only in total volume, standard similarity measures detect strong item associations — dubbed phantom correlation. The authors prove independent consumption under varying activity follows a Rasch model, derive how association depends on item popularity for each measure, and quantify phantom correlation's extent against real-world benchmarks.
More from Research
- Frontier models hit 73.3% on SpatialBench 10 months after release, near 75% ceiling — kenbwork · 2026-10-28
- Minimal PyTorch project implements transformer autograd by hand for learning custom backward passes — ifaposto · 2026-10-28
- Ben Goertzel's toy simulations probe the geometry of superintelligence attractor basins — bengoertzel · 2026-10-28
- ReGE: turning recorded agent execution experience into synthesized guidance for self-improvement — TIGER-Lab · 2026-10-28
- Statistician: the Brazil event wasn't 'impossible' — the distribution was wrong — doodlestein · 2026-10-28
- NVIDIA PhysicsNeMo tutorial: AI surrogate models for vehicle crash analysis — NVIDIA Developer · 2026-10-28