S2-CAR: New recommendation framework beats 13 baselines across 3 domains

Sequential recommendation systems predict user preferences from interaction histories, but existing models often fail when behavior patterns are complex and heterogeneous. The core issue is that user interests shift subtly over time, yet current approaches either treat the full sequence as uniform or rely on rigid time-window segmentation that misaligns with actual intent boundaries. This leads to cross-intent interference and over-reliance on short-term signals. To solve this, researchers introduce S2-CAR (Segmentation-Supervised Complexity-Adaptive Recommendation), which models user intent as a continuous latent energy state. It uses a Context-Aware Soft Temporal Point Process (Soft-TPP) to segment interaction histories based on natural decay of that energy, rather than fixed intervals. This allows intent boundaries to emerge dynamically. Then, a Segment-Count-Adaptive Multi-Intent Extraction module hierarchically aggregates these intent-coherent segments into a compact set of multi-interest representations, capturing both long-term and short-term preferences.

Extensive experiments on three representative public benchmark datasets spanning movie (e.g., MovieLens), e-commerce (e.g., Amazon), and gaming (e.g., Steam) domains, against 13 strong baselines, demonstrate that S2-CAR consistently outperforms state-of-the-art methods across all datasets and metrics. Notably, the energy-based segmentation module is plug-and-play: when integrated into existing sequential recommendation backbones, it yields consistent improvements. This work is available on arXiv (2606.25415) and marks a significant step forward in modeling complex user behavior for recommendation systems, potentially leading to more personalized and adaptive user experiences.