Date of Award

5-16-2024

Publication Type

Thesis

Degree Name

M.Sc.

Department

Computer Science

Keywords

Long-tail Distribution;Popularity Bias;Team Recommender Systems

Supervisor

Ziad Kobti

Supervisor

Kalyani Selvarajah

Creative Commons License

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

Abstract

In the field of collaborative work, effectively forming teams is crucial for achieving success. The challenge of constructing cohesive teams lies in selecting from a vast array of potential candidates, each with distinct skills, experiences, and personal attributes. Team recommender systems are designed to pinpoint the ideal combination of experts who collectively meet the skill requirements necessary to achieve a common goal. Recently, researchers have started to examine this problem through neural architectures that recommend the team of experts by learning a relationship between the skills and experts space. However, these models often exhibit a popularity bias, which refers to the tendency of the systems to recommend disproportionately more popular teams. We have introduced a dual transfer strategy to enhance team recommendation performance, which involves transferring knowledge from head instances to tail instances at both the model and instance levels. At the model level, the strategy creates a meta-mapping from few-shot to many-shot models, which indirectly improves data quality and enhances learning representations for teams that are less popular by implicitly augmenting data at the model level. The proposed dual knowledge transfer method at the instance level employs curriculum learning to bridge the gap between popular and not popular instances, ensuring a smooth transition of meta-mapping head teams to the tail ones. Our evaluation criteria is that we expect to improve team recommendation quality, particularly for teams that are in the tail of the distribution. We demonstrate how our proposed platform effectively addresses the issue of popularity bias prevalent in current team recommendation methodologies. Further more, we validate its effectiveness by comparing it with leading approaches, employing the DBLP dataset in our analysis.

Download

Share

COinS