Topic Modeling Enhanced Tripartite Graph for Recommendation Using Metapaths

In this paper, we propose a novel recommendation system approach by using the advantage of item representations with topic modeling. Our proposed model exploits topics as the sub-categories of the items and enriches the heterogeneous user-item graph by transforming it into user-item-topic tripartite graph. Unlike most of the existing methods, the sub-categories are constructed by using only item titles as features which makes it an easy-to-apply methodology that is applicable in real-world systems even if the items do not have description texts. We also utilize meta-paths to traverse the obtained tripartite graph and try to solve the sparsity problem in recommender systems by learning node embeddings using meta-path based Metapath2Vec algorithm. In order to observe the performance of our proposed model, we used Amazon datasets from two different categories and the results show that our model performs 3.8% better than even the closest baseline result when evaluated for both datasets according to the hit ratio. Further analysis verifies that incorporating sub-category information into graph structure yields an increase of recommendation accuracy especially when recommending long-tail items.