Vector Representation Learning of Skills for Collaborative Team Recommendation: A Comparative Study

Md Jamil Ahmed, Mahdis Saeedi, Hossein Fani

Abstract: Neural team recommendation models have utilized graph representation learning to achieve state-of-the-art performance in forming teams of experts whose success in completing complex tasks is almost surely guaranteed. Specifically, the proposed models frame the problem as an expert recommendation task for a set of required skills whose dense vector representations are transferred from a graph neural network on the collaboration graph. However, there is yet to be a systematic comparative study on the impact of (1) the collaboration graph structure, (2) the node representation learning technique, and (3) the architecture of the final neural recommender on the efficacy of recommended teams. In this paper, we establish a benchmark that includes two heterogeneous collaboration graphs and seven graph representation learning to learn dense vector representations of skills for variational and non-variational neural recommenders. Our experiments on two large-scale datasets from various domains, each with distinct distributions of skills within teams, in a host of classification and information retrieval metrics show that i) those graph neural networks that utilize attention on ii) heterogeneous collaboration graphs, including expert, team and skill nodes, consistently yield the best dense vector representations of skills for iii) neural team recommenders of different architectures iv) across datasets. The code to reproduce the experiments reported in this paper is available at https://github.com/fani-lab/OpeNTF/tree/wise24.