The School of Computer Science is pleased to present…
Dynamic Loss-Based Curriculum Learning on Neural Team Formation
MSc Thesis Proposal by:
Reza Barzegartorghabeh
Date: December 8th, 2023
Time: 11:00 am – 12:00 pm
Location: Toldo Lancer Centre, Room 204
Abstract:
Recommending collaborative teams of experts who, more likely than not, can solve complex tasks has firsthand effects on creating organizational performance and, hence, has been a surge of research interest in many disciplines. Existing neural team recommendation models have brought state-of-the-art efficacy while enhancing efficiency for the iterative and online learning procedure, and availability of training datasets; albeit their performance has yet to be improved for real-world applications with the large number of experts and disproportionate distribution of popular experts over successful teams. Specifically, such models frame the team recommendation problem as a multilabel Boolean classification task and assign a label to each expert, falling short of considering the difficulty of recommending hard nonpopular vs. easy popular experts. In this paper, we propose to leverage curriculum-based learning strategies to dynamically identify popular experts from nonpopular ones based on the model's loss during learning. We present two dynamic curricula to improve upon the existing neural models. Our proposed curricula offer a plug-and-play solution, obviating the need for changes in model architecture or the loss function. Our experiments on two large-scale datasets from different domains with varied distributions of teams over skills demonstrated consistent synergistic effects of our proposed curricula, evidencing their capacity to address the popularity bias.
Thesis Committee:
Internal Reader: Professor Luis Rueda
External Reader: Doctor Mohammad Hassanzadeh
Advisor: Doctor Hossein Fani
