In this thesis, we present a graph-based scalable and novel approach for the recommendation which doesn’t depend on the profile data for predictions. We evaluate our method on an extensive transaction dataset from the retail domain (700k transactions with 150k different items) and compare it to a baseline. The proposed approach relies on knowledge graph embeddings. During our evaluation, we have used two knowledge graph embedding algorithms. The suggested method first applies Pyke, a knowledge graph embedding algorithm with a close to linear runtime complexity. Later Pyke was replaced by a convolutional complex knowledge graph embedding algorithm, Conex. The evaluation results suggest that Conex fits better than Pyke to the approach.
Please look at the wiki page for detailed description