Link Prediction using metadata information

Performs scientific literature keyword-keyword co-occurrence prediction based on associated metadata based features.

Getting Started

Prerequisites

• NetworkX 2.2
• Keras
• theano
• statsmodel
• scipy
• numpy
• matplotlib
• pandas
• sklearn

Files

• link_prediction.ipynb - End to end link prediction (direct classification using feature - no forecast) experiment
• timeseries_forecast.ipynb - End to end link forecast-> classification experiment (first forecasting, then classification- the one used in this experiment)
• Link_Analysis.ipynb - Data generation regarding keyword network evolution and associated characteristics
• timeseries.ipynb - LSTM based Timeseries analysis of nodal degree
• graphs.py - Contains required functions to build, save and load graphs
• utils.py - Utility functions
• classification.py - Initial training-test set preparation, model training and evaluation
• feature_selection.py - Node level and edge level feature generation
• versions.py - Checks the versions of different packages

Dataset

• Apnea dataset - All keywords are atleast 3-degree keywords
• Apnea keyword lists - Keyword list with integer id
• Obesity dataset - All keywords are atleast 3-degree keywords
• Obesity keyword lists - Keyword list with integer id

Usage

1. Open link_prediction notebook. End-to-end link prediction experiment is done here (graph build, save, load -> training data prepare, save, load -> model training, save, evaluate -> result save, load -> figure generate, save)
2. Experimental analysis related to keyword network evolution is done in link_analysis notebook.
3. LSTM timeseries forecasting of top-3 central keywords nodal degree is done in timeseries notebook. Then ground truth vs predicted value graph is generated.

Authors

• Fahim Faisal
• Dr. Nazim Choudhury

License

MIT