Graph Neural Networks for Recommender Systems

This work was aimed to research initial transformations of user-item bipartite interactions graphs to improve quality metrics of graph neural networks for recommender systems' problems. It was set a task to create concrete methods for initial transformations of graphs of interactions based on fundamental graph properties and current known problems of recommender systems. Next it was needed to train graph neural networks using collected graphs. It was essential to compare methods using various quality metrics on static test graphs and perform comparative analysis of different properties of collected graphs. As a results of completing this work, there were proposed three methods that managed to increase some of the metrics without drastic changes among the rest metrics. Methods used knowledge about degree and betweenness centrality of nodes. Based on these properties it was created a bunch of latent edges or deleted some of the existing nodes. Transformed graphs were deeply analysed and some connections between concrete transformations and changing in quality metrics of outcome recommendations were revealed. Based on this work, some options of further researches in initial graph transformations were recommended. For example, it is recommended to stay out of quantiles and strong thresholds in favor of using knowledges about nodes' degree distribution. Another example is to use several methods consequently. To better understand the effects of the methods, it was proposed to perform online A/B tests to further reveal changes in user behaviour, because offline tests on static data are rather limited.