Date of Award

2-1-2025

Publication Type

Thesis

Degree Name

M.Sc.

Department

Computer Science

Keywords

Graph Neural Network; Neural Team Recommendation; OpeNTF; Social Information Retrieval

Supervisor

Hossein Fani

Rights

info:eu-repo/semantics/openAccess

Abstract

Neural team recommendation has leveraged graph representation learning to achieve state-of-the-art performance in recommending teams of experts whose success in completing complex tasks is almost surely guaranteed. The proposed models frame the problem as a Boolean multi-label classification, mapping a subset of required skills, whose dense vector representations are transferred from expert collaboration graphs by graph neural networks, to the sparse occurrence (multi-hot) vector representation of an optimum subset of experts using multilayer feedforward neural models. Such models, however, withhold two key shortcomings. (1) They learn the skill’s dense vectors in an unsupervised manner in a pre-training phase separately, oblivious to the supervised information about successful teams, resulting in suboptimal performance for the multi-label classification by the feedforward neural models. (2) Furthermore, they suffer from the curse of sparsity of the high-dimensional multi-hot vector representation of optimum experts in the output layer. In this paper, we propose an end-to-end graph neural network-based approach to the team recommendation problem, eschewing the unnecessary complications of the two-phase transfer learning and neural model fine-tuning. We reformulate the problem into a link prediction in a heterogenous expert collaboration graph based on the successful teams in a training set to directly and jointly learn dense vectors of skills and experts and recommend the optimum subset of experts as a team through predicting links between expert nodes and the nodes for the required subset of skills. Our experiments on two large-scale datasets from various domains, with distinct distributions of skills in teams, including dblp (computer science publications) and imdb (movies), showcase the superiority of our proposed method by about 24x in dblp and 29x in imdb against the baselines in a host of classification and information retrieval metrics. The code to reproduce the experiments reported in this paper is available at https://github.com/fani-lab/OpenGNTF/

Download

Share

COinS