Research On Knowledge-aware Recommendation System

With the exponential growth of information on the Internet,people have to face the problem of information overloading,which makes it difficult for users to select what really interests them from a large amount of content.To enhance user experience,researchers and engineers have introduced recommendation systems to target user interests and thus provide users with personalized recommendations.As the core of a recommendation system,the recommendation algorithm determines the final recommendation performance.A widely used recommendation method is collaborative filtering（CF）.However,CF-based methods usually suffer from data sparsity and cold-start problems.Although model-based CF methods can solve the problem of data sparsity to some extent,these methods usually fail to fully model the similarity among users or items.Another line of approaches to address the problems of sparse data and cold-start is to combine auxiliary information,and knowledge graph（KG）,as an effective source of auxiliary data,has received a lot of attention in academia and industry.To make full use of the semantic and connectivity information of entities and relations in the KG,the idea of graph neural networks has been introduced to learn KG.However,these methods fail to take full advantage of the neighborhood information of entities and ignore the importance of user interaction sequences on reflecting user preferences.Therefore,based on the above limitations,this paper carried out the following work:（1）Aiming at the problem that model-based CF methods fail to fully capture the similarity among users or items,a graph convolution collaborative filtering（GCCF）model is proposed.GCCF captures user interaction behavior similarity and item interaction behavior similarity based on user-item interaction graph and further construct user graph and item graph,and uses a simple and efficient graph convolution architecture to learn user representation and item representation on user graph and item graph,respectively.In addition,a highway gate mechanism is introduced to combine multi-layer graph convolutional networks to mine deeper similarity information of users and items.Based on the above design,the GCCF method can not only capture the similarity among users or items during the embedding process,but also reduce the impact of noise on user representations and item representations when aggregating similarity information.（2）Aiming at the problem that knowledge-aware recommendation systems based on graph neural networks do not fully utilize the neighborhood information of entities and ignore the importance of user interaction sequences on reflecting user preferences,a knowledge-aware hierarchical attention network（KHAN）is proposed.KHAN is equipped with a hierarchical attention network to carefully distinguish the importance of different neighboring nodes of an entity,achieving full utilization of neighboring information,and use the self-attention mechanism to capture the global dependencies among users’ interaction items to learn more informative user representations.（3）We have conducted extensive experiments on four real-world datasets for the proposed GCCF model and KHAN model,and evaluated them in two task scenarios about click-through prediction and Top-K recommendation,respectively.The experimental results show that the two models proposed in this paper significantly outperform the baselines,proving the effectiveness of the models.