Bilateral Intent-Enhanced Sequential Recommendation with Embedding Perturbation-Based Contrastive Learning

A research team led by Shanfan Zhang has introduced BIPCL (Bilateral Intent-enhanced, Embedding Perturbation-based Contrastive Learning), a novel AI framework designed to significantly improve sequential recommender systems. The core challenge it addresses is the accurate modeling of users' evolving preferences from their interaction history. While recent methods focus on latent intents, they often fail to leverage the collective intent signals shared across similar users and items, leading to isolated information and poor robustness. BIPCL tackles this with a bilateral intent-enhancement mechanism that creates shared intent prototypes for both users and items. These prototypes capture common behavioral semantics, which are then integrated into the learning process, alleviating information isolation and improving performance when supervision data is sparse.

Beyond intent modeling, BIPCL innovates in its use of contrastive learning—a technique for learning better representations by comparing similar and dissimilar data points. Instead of disrupting sequences with random augmentations, BIPCL injects bounded, direction-aware perturbations directly into the structural embeddings of items. This creates effective contrastive views while preserving crucial temporal and structural dependencies within a user's interaction sequence. The framework then enforces multi-level contrastive alignment across both interaction-level and the newly learned intent-level representations. According to the paper, extensive experiments on benchmark datasets demonstrate that BIPCL consistently outperforms current state-of-the-art baselines. Ablation studies further confirm the individual contribution of its bilateral intent enhancement and novel perturbation strategy, validating the design's effectiveness in building more accurate and robust recommendation models.