Asadullah Hill Galib

Asadullah Hill Galib

AI Research Scientist, AI for Business Intelligence (AI4BI) Center, TSMC Technology Inc.; PhD in Computer Science, Michigan State University. Specialization: Deep Learning, AI, Forecasting, Generative Modeling, Representation Learning, Time Series

Analyzing co-authorship network: Centrality Measure, Link Prediction, and Community Detection

[Github]

This is the course project of CSE 891: Computational Foundations in AI and Machine Learning (Fall 2020) at Michigan State University.

This project aims at analyzing a network of co-authorship relations between researchers (i.e., those in the top-300 highest hindex according to Google scholar and have collaborated at least once with each other). It includes network creation from XML data, data exploration, centrality measure using degree centrality, eigenvalue centrality, and local clustering coefficient. Also, it predicts missing links from the original network using Logistic Regression. Finally, it detects a set of connected components (communities) using a K-means clustering based community detection algorithm. The distinct parts of this project:

Network data creation

The goal is to create a co-authorship network from the given input files (DBLP XML Files) and store their relation in an output ASCII text file named coauthor.csv, which contains 2 comma-separated columns, (authorID1, authorID2), if the two authors have collaborated at least once in the past. The mapping from author name to author ID is given in the file authors.txt. Finally, I have visualized n co-authorship network with co-authorship links.

Data exploration (Centrality Measure)

  1. Compute the number of nodes (vertices) and links (edges) of the network
  2. Centrality Measure:
    1. Degree Centrality: Compute the degree (i.e., number of collaborators) associated with each author.
    2. Eigenvalue Centrality: Find the principal eigenvector of the adjacency matrix using power method.
    3. Local Clustering Coefficient: Compute the local clustering coefficient of each node in the graph.

The objective of link prediction in network analysis is to infer missing links or recommend new links in a network. For this step, tasks include:

  1. Removing some of the links in the original network to create a new, modified network.
  2. Using the modified network to train a logistic regression model.
  3. Applying the logistic regression model to predict the links that were removed in part 1.

Community Detection

The objective of community detection is to partition the network into a set of connected components (communities) such that nodes in the same community are more similar (connected) to each other compared to those belonging to different communities.