Karaveli Ifigeneia (Phd Candidate)

Thesis title: Graph based machine learning methods for providing personalized recommendations

Supervisor: Salampasis Michalis

Advisory Committee Members:

Konstantinos Diamantaras, Professor, Dept. of Information and Electronic Engineering, IHU
George Stalidis, Department of Organisation Management, Marketing and Tourism, IHU

Abstract:

One of the popular problems addressed with this new methodology is providing personalized recommendations. Recommender systems are tools for finding relevant information among ever increasing choices and have become very popular in the digital world. Users recommend items, movies, or any type of content they find interesting. These recommendations are typically based on user or item
characteristics, or users' past clicks, purchases, and interactions. The available data is best represented in a graph. GNNs can exploit both content-based information (user and object attributes) and graph structure (user-object interactions,) whereas traditional models can typically exploit only one of the two. The characteristic of a GNN is that it uses a form of neural messaging so that information is exchanged
between nodes and updated using neural networks (Gilmer et al., 2017). The model gathers messages from the neighbors of the graph, and in turn, the messages coming from those neighbors are based on information gathered from their respective neighborhoods, and so on. In GNNs, representations are created for all users and objects. For all users, we predict their preference object using the representations. Building embeddings is done through information propagation, also called neural message passing. The prediction of preferences can be done through cosine similarity. The GNN model consists of as many layers as we want. Each layer exchanges information between all immediate neighbors in the graph, and the number of layers determines how far the information propagates. The generated representations are used to predict the probability of a connection between two nodes.