Graph convolutional network based friendship recommender system for a social network

Abstract: This thesis aims to design and develop a deep neural network for friendship recommendations, and quantify its superior performance over a non-learned model for a given social network platform. The given social network consists of a large number of users over 120,000. It is important to characterize each user to offer relevant services on the platform such as friendship recommendations. Some users fill their profile information while some do not. Some users have made friends while some are just new to the platform and have no friends. A good recommender system should be able to tackle such realistic cases by utilizing the available (although limited in few cases) information about each user. It must be able to incorporate both the user profile information as well as the user friendship connection information. It is challenging for traditional non-learned models to incorporate both these set of information and recommend well. The field of deep learning has recently evidenced many innovations and advances and one of them is Graph Convolutional Networks (GCN). Lately, GCN-based models have proved their excellence in several recommendation tasks and shown state-of-the-art performance [1]. A Social network can be expressed as a graph with users as nodes and friendship connections as edges. Here, we examine the usage of a graph convolutional network to tackle the recommendation problem. Also, this thesis employs state-of-the-art pre-trained language model S-BERT for semantic representation of user textual data. This thesis investigates the viability of GCN-based models for friendship recommendations for individual social network users. In particular, a GCN model shall be developed based on PinSage architecture [1]. Then the GCN models shall be evaluated for three different setups. The obtained results show that the GCN-based model outperforms the non-learned model by a definitive margin. In particular, the findings show that a GCN model trained on a single city with rich user features yields the best results. This exhibits the versatility of GCN and corroborates its recent success in recommendation systems. This work demonstrates the rich characterization of each node in the graph (or user), which not only can be used to offer friendship recommendations but also to recommend other entities on the platform.