Side information aggregated convolutional neural network in news recommendation

Existing news recommendation models generally consist of a text feature extraction network and a recommendation network. News-related side information, such as category information, is not integrated into the text feature extraction process. Without fusing side information, there are differences in the optimization objectives between the text feature extraction network and the recommendation network. This paper proposes the structure of SIACNN (Side Information Aggregated CNN), which incorporates side information into text feature extraction through an attention mechanism. This narrows the gap between the optimization objectives of text feature extraction and the recommendation network, effectively improving the performance of news recommendation. Quantitatively, the integration of SIACNN into existing architectures resulted in average improvements of 0.011 in AUC and 0.012 in MRR, solidifying its effectiveness. SIACNN is used to replace the convolutional neural networks in several typical news recommendation networks, and experiments are conducted on the large-scale news dataset MIND (MIcrosoft News Dataset) collected from MSN (Microsoft News). The experiments demonstrate that SIACNN can enhance recommendation performance while possessing good generalization ability.