基于时间敏感的异构图相似神经网络的新闻推荐

Time-Sensitive Heterogeneous Graph Similarity Neural Network for News Recommendation

摘要: 目前网络上新闻信息呈现爆炸式增长趋势，为方便用户快速找到他们感兴趣的新闻，个性化新闻推荐变得愈发重要。目前主流的新闻推荐方法大多都是基于静态数据进行研究，忽略了可以间接反映用户对新闻兴趣程度的动态信息，如用户阅读新闻的时间等。为了解决此类问题，文中在GNewsRec模型的基础上提出了一种基于时间敏感的异构图相似神经网络模型(TSHGSN)。采用CNN学习新闻特征，加入用户阅读新闻时间权重并利用LSTM学习用户点击新闻的序列特征作为用户短期偏好。同时，构建了一个异构图，建模用户–新闻–主题关联，采用新的邻居采样方法聚合节点获取候选新闻特征表示和用户长期偏好。最后，将用户短期兴趣和长期兴趣与候选新闻分开进行相关性计算，旨在自适应的调整用户建模中短期兴趣和长期兴趣的重要性。实验数据表明，与GNewsRec模型相比，该模型在AUC指标上提高约4%。

Abstract: The explosive growth of news information on the internet has made personalized news recommendation increasingly important for helping users quickly find news articles of interest. Most mainstream news recommendation methods currently rely on static data, overlooking dynamic information that could indirectly reflect user interests in news, such as the time users spend reading news articles. To address this issue, this study proposes a time-sensitive heterogeneous graph similarity neural network model (TSHGSN) based on the GNewsRec model. The TSHGSN model incorporates user reading time weights into the learning of news features using CNN, and utilizes LSTM to learn the sequence features of user-clicked news as short-term user preferences. Additionally, it constructs a heterogeneous graph to model user-news-topic relationships and aggregates node representations of candidate news features and user long-term preferences using a novel neighbor sampling method. Finally, the model separates user short-term and long-term interests from candidate news for relevance calculation, with the aim of adaptively adjusting the importance of short-term and long-term interests in user modeling. Experimental results show that compared to the GNewsRec model, this model achieves approximately a 4% improvement in the AUC metric.