Date of Award

2023

Publication Type

Thesis

Degree Name

M.A.Sc.

Department

Computer Science

First Advisor

H.Fani

Second Advisor

R.Gras

Third Advisor

S.Samet

Keywords

Machine learning, Neural networks, Negative sampling, Bayesian neural model, Neural team formation

Rights

info:eu-repo/semantics/openAccess

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Abstract

Predicting future successful teams of experts who can synergistically work in concert with each other and en masse cover a set of required skills of a degree necessary for the achievement of the desired outcome is challenging due to several reasons, including 1) the magnitude of the pool of plausible expert candidates with diverse backgrounds and skills, and 2) the drift and variability of collaborative ties of experts and their level of expertise in each area in time. Prior works in team formation have neglected the fact that experts’ skill, interests, and collaborative ties change over time. We can categorize previous works in team formation based on their method of optimization: 1) search-based, where the search for the optimum team is carried over all the subgraphs of expert networks or via integer programming, however, these works overlooked the temporal nature of human collaborations. 2) learning-based, where machine learning approaches are used to learn the distributions of experts and skills in the context of successful teams in the history to predict almost surely successful teams in the future. However, they also fail to recognize the possible drift and variability of experts’ skills, interest, and collaborative ties in time and its impact on the prediction of future successful teams. Moreover, neural models are prone to overfitting when training data suffers from the long-tail phenomenon, i.e., few experts have a lot of successful collaborations and the majority have participated sparingly. To overcome the aforementioned problems, i) we propose a streaming scenario training strategy for neural models to help the model in the prediction of future successful teams of experts, where instead of shuffling our datasets, we train the models in an orderly manner, to grasp the changes in experts’ skills, interests, and collaborations, and ii) we propose an optimization objective that leverages both successful and virtually unsuccessful teams via various negative sampling heuristics, and iii) we conduct experiments on four large-scale benchmark datasets with varying distribution of skills and members namely, dblp, imdb, uspt, and github. Finally, we empirically demonstrate how our proposed objective functions and training method, outperform the state-of-the-art approaches in terms of effectiveness and efficiency.

Share

COinS