Latent Intent Anchoring for Robust Sequential Recommendation
Abstract
User behavior sequences are often driven by stable, high-level intents. Existing intent-aware sequential recommendation methods typically extract intents from derived sequence representations to serve as post-hoc supervisory signals. Although effective, this post-hoc paradigm prevents intents from explicitly participating in the semantic shaping of sequential representations. Without a structural high-level semantic anchor, the sequence representations become vulnerable to local item transitions and user behavioral noise. To overcome this challenge, we propose IASRec, an Intent-Anchored Sequential Recommendation framework that shifts user intents from post-hoc supervisors to architectural anchors. IASRec consists of three synergistic modules: (1) a Latent Intent Distiller (LID) to extract multi-faceted latent intents from interaction sequences; (2) an Intent-Anchoring Encoder (IAE) that infuses intent semantics into item representations to guide temporal dependency modeling; and (3) an Intent Consistency Regularizer (ICR) that ensures robustness across different intent views. Extensive experiments on three real-world datasets demonstrate that IASRec achieves an average 7.7% improvement over state-of-the-art baselines and exhibits superior robustness, with markedly better resilience to 20% behavioral noise compared to competing methods. The source code is available at https://github.com/x-labs-xyz/iasrec.