Joint Knowledge Pruning and Recurrent Graph Convolution for News Recommendation

Recently, exploiting a knowledge graph (KG) to enrich the semantic representation of a news article have been proven to be effective for news recommendation. These solutions focus on the representation learning for news articles with additional information in the knowledge graph, where the user representations are mainly derived based on these news representations later. However, different users would hold different interests on the same news article. In other words, directly identifying the entities relevant to the user's interest and deriving the resultant user representation could enable a better news recommendation and explanation.

To this end, in this paper, we propose a novel knowledge pruning based recurrent graph convolutional network (named Kopra) for news recommendation. Instead of extracting relevant entities for a news article from KG, Kopra is devised to identify the relevant entities from both a user's clicked history and a KG to derive the user representation. We firstly form an initial entity graph (namely interest graph) with seed entities extracted from news titles and abstracts. Then, a joint knowledge pruning and recurrent graph convolution (RGC) mechanism is introduced to augment each seed entity with relevant entities from KG in a recurrent manner. That is, the entities in the neighborhood of each seed entity inside KG but irrelevant to the user's interest are pruned from the augmentation. With this pruning and graph convolution process in a recurrent manner, we can derive the user's both long-and short-term representations based on her click history within a long and short time period respectively. At last, we introduce a max-pooling predictor over the long-and short-term user representations and the seed entities in the candidate news to calculate the ranking score for recommendation. The experimental results over two real-world datasets in two different languages suggest that the proposed Kopra