Social Network Analysis: Author Success Prediction

Files

• soc_net_analysis.ipynb: Jupyter Notebook containing the full implementation of network analysis and predictive modeling.
• sna_report.pdf: Final report detailing the methodology, experimental design, and results.

Overview

This project investigates the relationship between a researcher's position in a coauthorship network and their likelihood of publishing in top-tier venues. It is inspired by the study by Sarıgöl et al. (2014), which used citation counts to predict academic success through social network features. This project redefines success based on publication venue prestige.

Objectives

• Construct a time-evolving coauthorship graph from the DBLP dataset (2019–2025).
• Define a binary success label: papers published in top-tier venues (e.g., ICLR, NeurIPS, CVPR) are marked as successful.
• Extract author-level network features (Betweenness Centrality, K-Core Number) using the Networkit library.
• Aggregate features at the paper level and train machine learning models to predict publication success.

Methods

1) Network Construction

• Nodes represent authors, and edges represent coauthorship.
• Networks are created using 2-year sliding windows to capture temporal evolution.
• Isolated nodes and self-loops are removed; the largest connected component is retained for analysis.

2) Feature Extraction

• Centrality metrics are computed per author and aggregated per paper (e.g., maximum betweenness among coauthors).
• Features used:
  • Betweenness Centrality
  • K-Core Number

3) ML Models

• Random Forest Classifier: Robust to small datasets, serves as a baseline.
• Multi-Layer Perceptron (MLP): Includes feature scaling and SMOTE oversampling to address class imbalance.

Models are trained and tested on datasets of increasing size: 100K, 200K, 500K, and 1M papers.

References

• Sarıgöl, E., et al. Predicting Scientific Success Based on Coauthorship Networks. EPJ Data Science, 2014.