Node embedding algorithms in product recommendation systems

Sammanfattning

Product recommendation systems are used for recommending products or product- groups to users or user-groups. With the rise of big data, these systems are in- creasingly being built on machine learning algorithms that use the collection of data to train and produce recommendations. How to represent the collected data for the prediction algorithm in a meaningful way can be a complex issue. Node embedding algorithms is utilised for latent representation learning of graphs for downstream prediction task. A good node embedding should capture the graph topology, the node-to-node relationships, and all the relevant information that is needed to make good prediction. In this paper we propose a pipeline for testing the usage of node embedding algorithms to learn the representation of real-world data that is then used to do link prediction on. Based on the experiments run on the pipeline we formulate the following conclusions. We confirm that using node2vec to formulate latent representations of bipartite graph data for link prediction has the potential to produce good results when working with a temporally static model. The node embedding algorithm captures enough information of the underlying topology of the graph to be able to predict missing links within the graph. We show that the addition of hyperparameter optimisation techniques can be very costly within the constrains set by product recommendation systems, and that only viable version of these hyperparameter optimisation techniques would require a lot of previous knowledge about the properties of the input dataset in terms of size and completeness.