A multi-model recurrent knowledge graph embedding for contextual recommendations

Abstract

Recommendation Systems are a key part of every modern information system. From social media platforms to restaurants and hotel recommendations, there is a huge disposal of multi-context data that could be used for training and creating a recommender. Knowledge Graphs (KGs) are a great way to incorporate this kind of information, by using paths and meta-paths to retrieve the needed information for recommending something to the user. In this thesis, we introduce MRKGEC, a meta-path, query-based, approach to mining paths from a graph database that stores our KG and creates an embedding approach that learns the representations of meta-paths and recommends items to the user on this basis. Our system uses a number of LSTM neural network models to encode the meta-path semantics between a user-item pair, based on the length of the mined path. We have also included a Multi-head Attention module as an attention mechanism, in addition to a pooling and a recommendation layer. Evaluating our system on the Yelp dataset shows that MRKGEC is on par with modern recommendation systems. Lastly, we exhibit how our system is able to mine semantic paths based on the needed context and compare different contextual versions of our system against themselves while showing that contextual modelling outperforms random testing on precision and MRR metrics.