基于会话的双视角社交推荐模型

Dual-View Model for Session-Based Social Recommendation

摘要: 现有的基于会话的社交推荐模型单一的通过图视角学习会话的嵌入表示并做出预测，而忽略了项目之间的顺序依赖性。针对该问题，提出了一种基于会话的双视角社交推荐模型(Dual-view approach for session-based social recommendation, DVSSR)。DVSSR分别从图视角以及行为序列视角学习会话嵌入。在会话图视角中，通过图神经网络传播节点之间的特征信息，通过注意力机制聚合会话嵌入表示；在行为序列视角中，使用自适应旋转位置编码学习项目之间的相对位置关系，通过多头自注意力机制学习序列嵌入表示，将两种视角下的嵌入表示融合并做出推荐。然后，为了确保双重视角下学习到用户偏好的一致性，设计对比学习模块拉近两种视角下学习嵌入表示，并通过破坏生成增强视图与原视图对比学习提高模型的鲁棒性和泛化能力。最后，在三个大型公开数据集上的实验结果验证了DVSSR的有效性和优越性。

Abstract: Session-based social recommendation aims to predict the next most likely item of interaction for the current session based on the user’s social network and historical session information. Existing session-based social recommendation methods predominantly learn session embeddings from a single graph perspective for making predictions, thereby neglecting the sequential dependencies among items. To address this issue, we propose a Dual-View approach for Session-Based Social Recommendation (DVSSR). DVSSR learns user and item embeddings through a user-item heterogeneous graph and learns session embeddings from both graph and behavioral sequence perspectives. In the session graph perspective, it propagates feature information between nodes using a graph neural network and aggregates session embeddings via an attention mechanism. In the behavioral sequence perspective, it employs an adaptive rotary position encoder to learn the relative positional relationships among items and captures the sequential dependencies and sequence embeddings using a multi-head self-attention mechanism. The results from these two perspectives are then integrated with user embeddings to form the final session embedding for making predictions. To ensure consistency in user preference learning across the dual perspectives, a contrastive learning module is designed to align the learned embeddings from both views, and the method’s robustness and generalization capability are enhanced through contrastive learning between augmented and original views. Finally, experimental results on three large public datasets validate the effectiveness and superiority of DVSSR.