Context-aware recommender systems with graph convolutional embeddings (CARS-GCE)

Abstract

[en] The amount of online service providers is increasing every year, including multimedia streaming services and online shops. These services show a great interest in accurately recommending more products and more content to their costumers, as this strategy clearly encourages their clients to purchase or consume more items provided by them. Recommender Systems are a useful tool that automatizes the task of predicting the preferences of the users of a service in order to recommend them items that will match their taste. Research on this area generally seeks for ways to improve the performance of the mathematical models these systems are based on in order to obtain better recommendations as result. In this work, our main goal is to understand some traditional models used for recommendation and extend them so that they can detect complex patterns in the ratings given by users to items, capturing high-order interactions between features. Also, we aim to adapt them as Context-Aware Recommender Systems, which also take into account information about the context in which a user consumes a given item while computing their predictions. First, the recommendation problem and Recommender Systems will be clearly defined and then, two traditional models will be introduced: Matrix Factorization and Factorization Machines. These both models are related to the concept of Embedding, which will also be detailed. It will be explained that these models present limitations that prevent them from capturing high-order interactions between features. We aim to give the models the ability to capture these high-order interactions by using Graph Convolutional Networks (GCN) instead of the Embedding Layer. GCNs allow us to approach the recommendation problem as a graph link prediction problem, called Graph Convolutional Matrix Completion. GCNs encode the information of each feature in the graph and aggregates to it the correlated knowledge from neighboring features in the graph. Then, the graph structure will be adapted so that context information can be included in it. Also, the models will be fed with item metadata, formatted as side-information. Finally, we will detail the data used to train the model, how this data is treated and how the model is configured. In order to fairly compare the results obtained by each model, each one of their optimal settings will be calculated through Bayesian Optimization. Afterwards, we will expose and analyze the results. To conclude, it should be remarked that the inclusion of Graph Convolutional Networks with context information in the model implementation has a great positive impact on the results. Also, working with context in a traditional embedding structure may be benefitial only for specific models. The addition of item metadata shows different behavior depending on the metadata added and the model that is being evaluated. In future work, we plan to check whether adding item metadata into the graph structure may have better results than including it as side-information. Also, we would aspire to extend the Bayesian Optimization to more model parameters and compare the model performances with different data representations and loss functions, among others.