新闻推荐系统中的边信息融合卷积神经网络

Side Information Aggregated Convolutional Neural Network in News Recommendation

摘要:

现有的新闻推荐模型一般由文本特征提取网络和推荐网络两部分组成。新闻相关的边信息（如类别信息）并没有作用在文本特征提取过程中。在未融合边信息的情况下，文本特征提取网络和推荐网络两部分的优化目标是有差异的。提出SIACNN（Side Information Aggregated CNN）的结构，它通过注意力机制的方式，将边信息结合到文本特征提取中，缩小了文本特征提取和推荐网络之间优化目标的差异，有效提升了新闻推荐的效果。将SIACNN替换多个典型新闻推荐网络中的卷积神经网络，并利用MSN（微软新闻）采集的大型新闻数据集MIND（MIcrosoft News Dataset）来进行实验，通过实验证明了SIACNN能提高推荐效果，并同时具有泛化性。

Abstract:

Existing news recommendation models generally consist of the text feature extraction network and the recommendation network. News-related side information， such as category， is not fused into the text feature extraction network. Without fusing it， there are differences between the optimization targets of the text feature extraction network and the recommendation network. In this paper， a general SIACNN （side information aggregated CNN） layer is proposed. The SIACNN layer fuses the side information into the text feature through the attention mechanism， which bridges the gap between text feature extraction and recommendation tasks and improves the effectiveness of the recommendation. CNNs are replaced in many state-of-the-art models which used CNNs to extract the text feature with the SIACNN and several experiments are conducted in a large real-world news recommendation dataset MIND（MIcrosoft News Dataset） collected from MSN（MicroSoft News）. The recommendation effectiveness and generality of SIACNN are verified by several experiments.