Abstract
Sequential recommendation systems, despite their progress, still struggle with two core challenges: perceiving complex temporal dynamics in user behavior and modeling users' diverse interests. To address these limitations, this paper proposes FASR, a novel Future-Aware and Smooth Rhythm-aware Time-Gated GRU Recommender. At the core of our model is the innovative SRT-GRU encoder, which features a dual temporal gating mechanism composed of a rhythm-aware short-term gate and a smooth rhythm-aware session gate. This allows SRT-GRU to deeply understand behavior intervals, density, and rhythm, encoding this understanding into context-aware states enriched with model “confidence”. Building on this, we employ a dualstream architecture derived from users' similar interactions. A Historical Stream encodes the user's own history, while a Future Stream encodes “future behavior” sequences mined from neighbors, achieving “future-awareness”. Subsequently, a Capsule Network with dynamic routing acts on these timeaware contextual sequences to disentangle diverse historical interests and future trends. Finally, a target-aware readout and fusion mechanism generates a final user representation that simultaneously captures past preferences, potential future trends, and multiple interests for recommendation. Extensive quantitative and qualitative experiments on three real-world datasets (MovieLens-1M, Movies&TV, Movies&T, VOD) demonstrate that our model significantly outperforms state-of-the-art sequential recommenders on various metrics, while also providing complementary and novel recommendations. Ablation studies further confirm the effectiveness of each temporal component within our proposed SRT-GRU encoder.