融合行为交互图的兴趣感知新闻推荐算法

Interest-aware news recommendation algorithm fused with behavioral interaction graphs

个性化新闻推荐的关键是候选新闻和用户兴趣的精准匹配,现有基于顺序模型的方法通过建模行为序列的单向交互仅能捕获单一的用户兴趣,而基于图的方法通常忽略了用户行为序列内部的高阶转换关系。针对上述问题,提出了融合行为交互图的兴趣感知新闻推荐模型,以全局和局部的角度建模用户动态兴趣。该模型结合知识图谱和深度预训练网络以多视图学习方式提取新闻深层语义信息,采用融合上下文位置信息的线性自注意力机制捕获局部用户兴趣。此外,将用户行为序列构建为有向交互图,使用门控图神经网络递归地聚合邻域信息捕获序列间的高阶转换关系,从而挖掘全局用户兴趣。在2个公开数据集上的实验结果表明,本文提出的模型在各个指标上均优于基线模型,并且有效提高了新闻推荐效果。

The key to personalized news recommendation is the accurate matching between candidate news and user interests.Existing methods based on sequential models can only capture a single user interest by modeling the one-way interaction of behavior sequences,while graph-based methods often ignore the higher-order transition relationships within user behavior sequences.To address these issues,a news recommendation model with behavior interaction graph fusion is proposed,which models user dynamic interests from both global and local perspectives.The model combines knowledge graphs and deep pre-training networks to extract deep semantic information from news using a multi-view learning approach.The linear self-attention mechanism with fused contextual position information is used to capture local user interests.In addition,the user behavior sequence is constructed as a directed interaction graph,and a gated graph neural network is used to recursively aggregate neighborhood information to capture higher-order transition relationships between sequences,thereby mining global user interests.Experimental results on two public datasets show that the proposed model outperforms baseline models in all indicators and effectively improves news recommendation performance.