Link prediction on CV graphs : a temporal graph neural network approach

This thesis studies one of the recent advancements on the graph and its application in terms of the power of predictability that they can endow. More clearly, Machine Learning on Graphs has recently drawn much attention as it has enabled us to perform Graph Classification, Node Recommendation, Node Classification, Node Clustering, and link prediction. Following an agreement with our industrial partner, Euranova, this thesis was assigned to scrutinize machine learning on dynamic networks and examine the popular Temporal Graph Network --- a deep learning framework for dynamic graphs---, aka TGN, with a challenging curriculum-vita dataset with thousand applicants provided by Euranova. In case we mine the dataset and obtain a dynamic bipartite CV graph consisting of thousand nodes of distinct types (applicants and roles), the aim is to see how sufficiently TGN can help build a role predictor that could also be deemed a role recommender. This curiosity comes from the fact that people may change their roles over time as they get more experienced, educated, or seniority. For example, a junior data scientist will become a senior data scientist after several years, or a developer may become Development Lead as s/he acquires sufficient knowledge and experience. Thus, the time dimension defines and captures the evolution, and avoiding time in any models predicting the roles or jobs disregards this verity. In light of this reality, this thesis presents a role predictor and top-k recommendation system based on the provided CV Data and TGN model that can record the evolutionary patterns and recommend subsequent roles for applicants based on their experiences and skills. The experimental findings demonstrate that the proposed model is able to recommend k roles out of thousands of available roles such that the predicted roles closely resemble the ground-truth roles.