This repository provides the model weights and source code for the paper LISRec: Modeling User Preferences with Learned Item Shortcuts for Sequential Recommendation.

LISRec is a novel framework for sequential recommendation that addresses noisy user-item interaction histories by explicitly capturing stable preferences. It achieves this by extracting personalized semantic shortcuts from historical interactions, leveraging task-agnostic semantic representations to assess item similarities, and constructing a personalized semantic graph. By identifying the maximal semantic connectivity subset, LISRec selects representative items as semantic shortcuts to guide user preference modeling, filtering out irrelevant actions while preserving genuine interests. LISRec demonstrates strong generalization capabilities by enhancing both item ID-based and text-based recommendation models, achieving a 13% improvement over baseline recommendation models on Yelp and Amazon Product datasets.

Find the code and further details on the official GitHub repository.

LISRec addresses the challenge of noisy data in sequential recommendation by constructing a user-interacted item graph. It leverages item similarities derived from their text representations to extract the maximum connected subgraph, effectively denoising the items a user has interacted with.

The LISRec codebase requires several Python packages and OpenMatch. For detailed instructions on setting up the environment, downloading pretrained T5 weights, and installing OpenMatch, please refer to the official GitHub repository.

This work builds upon and acknowledges:

If you find this work useful, please cite our paper:

For questions, suggestions, or bug reports, please contact: xinhaidong@stumail.neu.edu.cn