LoopRec varies recursion depth per item in sequential recommenders
Original title: Think Harder Where It Matters: Item-Level Adaptive Recursion for Sequential Recommendation
A shared block and per-item exit gate beat matched fixed-depth models on public benchmarks across four backbone families.
The paper argues sequential recommenders waste a fixed compute budget because most history items converge early while a minority needs full depth. LoopRec reuses one weight-shared block and lets a lightweight gate exit or continue independently for each item, without depth labels. On public benchmarks with four backbone families, it outperforms matched fixed-depth models. Lower-frequency items and more recent interactions receive deeper computation. Code is at github.com/Elpmis117/LoopRec.
- Most items converge after shallow processing; a minority needs full depth.
- Oracle non-uniform depth outranks a fixed allocation.
- The gate makes a binary exit-or-continue choice at each recursion boundary.
- Training does not use auxiliary depth labels.
From the original article
1 min read at arxiv.orgExisting sequential recommenders usually allocate a fixed computational budget to every item in a user's interaction history, regardless of the item's contribution to the final prediction. We challenge this de facto design with two diagnostic findings. (1) Non-uniform computational need: items within the same sequence vary drastically in how much computation they require.
By Zhanshuo Cao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.