Improving Dynamic Recommendation using Network Embedding for Context Inference

Network embedding, which is a method to learn low-dimensional latent representations of nodes in networks, can be effectively utilized to infer contexts in the context-aware recommender domain. One of the fundamental challenges of network embedding is how to effectively and efficiently learn embeddings from dynamic networks, whose nodes and edges change over time. Network embedding approaches designed for static networks are infeasible to use with dynamic networks for reasons of scalability. The use of network embedding for inferring contexts in the incremental recommender task poses two fundamental challenges: (1) efficiently inferring contextual information that changes over time; and (2) integrating learned contextual features with a recommender technique that can be updated incrementally. To address these challenges, we present a neural recommender approach that models user interactions in the dynamic setting. Furthermore, we introduce a novel dynamic network embedding method based on an efficient neighborhood sampling technique, which employs a temporally biased form of random walk. We have successfully applied our approach to Point-Of-Interest recommendation domain by improving efficiency in context inference and quality of recommendations.

I. Introduction

In contrast to conventional recommender systems, which only utilize information about users and items, Context-Aware Recommender Systems (CARSs) incorporate contextual information and make the recommendation problem multidimensional. Producing hand-engineered features leads to high computational costs. Hence, it is important to efficiently infer contextual features from dynamic recommender environments. Network embedding, which is a promising technique for learning latent representations of nodes based on the structural properties of networks [1], has been used to infer contextual features for recommendations that are based on the static setting [2]. However, we identified two main limitations of current work that integrated network embedding as a context inferring mechanism with the recommender task.