Session-based recommendation aims to predict the next item based on the user–item interactions within the current session. Many existing methods adopt discriminative approaches to learn specific preference representations, while few methods introduce generative approaches to learn underlying preference distributions, failing to handle limited and noisy interactions effectively. Moreover, naive implementations of generative models face a trade-off between effectiveness and efficiency, limiting their practical utility. To address these challenges, we propose Dimos, a dual-branch framework comprising an exploring branch and an exploiting branch, which leverage diffusion models and attention networks to capture implicit and explicit preferences, respectively. At the core of Dimos is Bi-MaKAN, a novel backbone architecture featuring a pair of parameter-sharing bidirectional Mamba blocks and a Kolmogorov–Arnold network-based feature fusion layer, designed to enhance both performance and efficiency. To further improve generalization and reduce overfitting, we unify the sequential state spaces of both branches. Additionally, we introduce a linearly weighted fusion mechanism that integrates preference representations from both branches, enabling flexible adjustment of implicit and explicit preference contributions during training and inference. Extensive experiments on three real-world benchmark datasets demonstrate the superiority of Dimos, achieving up to 2.79% improvement in Recall, 3.09% in Mean Reciprocal Rank (MRR), and 3.00% in Normalized Discounted Cumulative Gain (NDCG) over state-of-the-art baselines. Efficiency evaluations show substantial gains, with reductions of 94.32% in Graphics Processing Unit (GPU) memory usage, 66.81% in training time, and 98.80% in inference time. In-depth analyses reveal a collaborative effect between the two branches during both training and inference, with dataset scale modulating their relative importance.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号