Relevant Article Recommendation by Learning Heterogeneous Network Embedding using GNN

Heterogeneous networks are powerful tools for representing complex multidimensional data. One such example is the bibliographic information network with diverse types of nodes and multi-relational edges. Recommendation on such complex data is a challenging task. In recent years, Graph Neural Networks(GNN) have gained significant attention due to their ability to learn representations by encoding complex topological structures and node attributes. Existing approaches learn embeddings by representing the heterogeneous network as layers of homogeneous networks. The embeddings learned from homogeneous layers are aggregated to represent the heterogeneous nodes, thus ignoring the interdependent information among the different types of nodes. We propose a deep learning approach for Article Recommendation using GNN(AR-GNN) in which heterogeneous information is captured at every node from the diverse neighborhood. In addition to giving the entire heterogeneous network as input to the GNN, we include a feature matrix by defining 3 types of heterogeneous node scores and a meaningful keyword representation using a network-based approach. We define a distance-based custom loss function to train the proposed models. We perform extensive experimentation on two heterogeneous networks DBLP V10, and DBLP V11. Instead of predicting the exact bibliography of a paper, we argue that retrieving papers of high quality and relevance is more important. Towards this end, we propose a quality metric and use it to compare the results of the proposed models with the existing approaches. The proposed approach clearly gives higher quality recommendations than the baseline methods. The predictions are also matched with the original bibliographies and the results are on par with the second best. The results obtained are satisfactory, as relatively, we use very limited meta-data information for this purpose.