Date of Award

5-21-2020

Publication Type

Doctoral Thesis

Degree Name

Ph.D.

Department

Computer Science

Keywords

Cultural Algorithms, Deep Learnings, Dynamic Networks, Link Predictions, Social Network Analysis, Team Formations

Supervisor

Ziad Kobti

Supervisor

Mehdi Kargar

Rights

info:eu-repo/semantics/openAccess

Abstract

Team Formation Problem (TFP) in Social Networks (SN) is to collect the group of individuals who match the requirements of given tasks under some constraints. It has several applications, including academic collaborations, healthcare, and human resource management. These types of problems are highly challenging because each individual has his or her own demands and objectives that might conflict with team objectives. The major contribution of this dissertation is to model a computational framework to discover teams of experts in various applications and predict the potential for collaboration in the future from a given SN. Inspired by an evolutionary search technique using a higher-order cultural evolution, a framework is proposed using Knowledge-Based Cultural Algorithms to identify teams from co-authorship and industrial settings. This model reduces the search domain while guiding the search direction by extracting situational knowledge and updating it in each evolution. Motivated from the above results, this research examines the palliative care multidisciplinary networks to identify and measure the performance of the optimal team of care providers in a highly dynamic and unbalanced SN of volunteer, community, and professional caregivers. Thereafter, a visualization framework is designed to explore and monitor the evolution in the structure of the care networks. It helps to identify isolated patients, imbalanced resource allocation, and uneven service distribution in the network. This contribution is recognized by Hospice and the Windsor Essex Compassion Care Community in partnership with the Faculty of Nursing. In each setting, several cost functions are attempted to measure the performance of the teams. To support this study, the temporal nature of two important evaluation metrics is analyzed in Dynamic Social Networks (DSN): dynamic communication cost and dynamic expertise level. Afterward, a novel generic framework for TFP is designed by incorporating essential cost functions, including the above dynamic cost functions. The Multi-Objective Cultural Algorithms (MOCA) is used for this purpose. In each generation, it keeps track of the best solutions and enhances exploration by driving mutation direction towards unexplored areas. The experimental results reach closest to the exact algorithm and outperform well-known searching methods. Subsequently, this research focuses on predicting suitable members for the teams in the future, which is typically a real-time application of Link Prediction. Learning temporal behavior of each vertex in a given DSN can be used to decide the future connections of the individual with the teams. A probability function is introduced based on the activeness of the individual. To quantify the activeness score, this study examines each vertex as to how actively it interacts with new and existing vertices in DSN. It incorporates two more objective functions: the weighted shortest distance and the weighted common neighbor index. Because it is technically a classification problem, deep learning methods have been observed as the most effective solution. The model is trained and tested with Multilayer Perceptron. The AUC achieves above 93%. Besides this, analyzing common neighbors with any two vertices, which are expected to connect, have a high impact on predicting the links. A new method is introduced that extracts subgraph of common neighbors and examines features of each vertex in the subgraph to predict the future links. The sequence of subgraphs' adjacency matrices of DSN can be ordered temporally and treated as a video. It is tested with Convolutional Neural Networks and Long Short Term Memory Networks for the prediction. The obtained results are compared against heuristic and state-of-the-art methods, where the results reach above 96% of AUC. In conclusion, the knowledge-based evolutionary approach performs well in searching through SN and recommending effective teams of experts to complete given tasks successfully in terms of time and accuracy. However, it does not support the prediction problem. Deep learning methods, however, perform well in predicting the future collaboration of the teams.

Download

Share

COinS