Author: Naw Safrin Sattar (University of New Orleans)
Advisor: Shaikh Arifuzzaman (University of New Orleans)
Abstract: Parallel computing plays a crucial role in processing large-scale graph data. Complex network analysis is an exciting area of research for many applications in different scientific domains, e.g., sociology, biology, online media, recommendation systems, and many more. Machine/Deep learning plays a significant role in working with big data in modern era. We work on a well-known graph problem, community detection (CD). We design parallel algorithms for Louvain method for static networks and show around 12-fold speedup. We also detect temporal communities in dynamic networks representing social/brain/communication/citation networks in a more concrete way. We present a shared-memory parallel algorithm for CD in dynamic graphs using permanence, a vertex-based metric. We also show the change of communities in different time phases computing several graph metrics based on their temporal definition. We present a scalable method for CD based on Graph Convolutional Network (GCN) via semi-supervised node classification using PyTorch with CUDA on GPU environment (4x performance gain). Our model achieves up to 86.9% accuracy and 0.85 F1 Score on different real-world datasets from diverse domains. To extend our work on deep learning, we provide a scalable solution to the Sparse Deep Neural Network (DNN) Challenge by designing data parallel Sparse DNN using TensorFlow on GPU (4.7x speedup). We include the applications of webspam detection from webgraphs (billions of edges) and sentiment analysis on social network, Twitter (1.2 million tweets) to reveal insights about COVID-19 vaccination awareness among the public to portray the importance of graph mining in our daily activities.
Thesis Canvas: pdf