Nonlinear Diffusion for Community Detection and Semi-Supervised Learning

The code used in Rania Ibrahim and David F. Gleich. "Nonlinear Diffusion for Community Detection and Semi-Supervised Learning" paper, accepted at The Web Conference (formally WWW), 2019.

Synthetic Data Experiments

You can find the code for LFR synthetic data experiments in LFR_Figures folder. To generate the conductance and F1 results, which are reported in figure 2 in the paper, run run_on_LFR_data.m

Community Detection Experiments

You can find the code for community detection experiments in Community_Detection_Experiments folder. Inside this folder there are three subfolders:

• community_detection: contains the experiments for nonlinear diffusion using power function.
• community_detection_tanh: contains the experiments for nonlinear diffusion using tanh function.
• community_detection_p_lap: contains the experiments for nonlinear diffusion using p-Laplacian.

To generate the results in table 1 of the paper: First download the datasets from SNAP website, add them folder inside a folder called "data" and then inside each subfolder run:

• run_com-dblp.jl: which produces the DBLP results.
• run_com-amazon.jl: which produces the Amazon results.
• run_com-youtube.jl: which produces the Youtube results.
• run_com-lj: which produces the Live Journal results.

Graph-based Semi-supervised Learning Experiments

You can find the code for graph-based semi-supervised learning experiment inside the folder Graph-based_Semi-supervised Learning_Experiment. Inside this folder, there are three subfolders:

• Nonlinear Diffusion with Growth or Decay: contains the code for running nonlinear diffusion using power function or tanh function.
• Nonlinear Diffusion via Nonlinear Transfer: contains the code for running nonlinear diffusion using p-Laplacian.
• Stability_Experiment: contains the code to generate figure 3 in the paper, where we vary the parameters for power function and tanh function and show that the classification accuracy remains high for many choices.

To generate the results in table 3 in the paper, inside folder Nonlinear Diffusion with Growth or Decay run run_cora.sh to get the results for the cora dataset, run run_citeseer.sh to get the results on citeseer dataset and run run_pubmed.sh to get the results on pubmed dataset.

To generate the results in figure 3, run run_stability_experiment.sh

FASHION MNIST Experiments

You can find the code for FASHION MNIST experiments inside the folder Fashion_MNIST_Experiments. Inside the folder, you can find:

• Visualization Folder: contains the code to generate figure 1 and figure 4 in the paper. In order to generate the figures, you need to:
  • Download FASHION_MNIST dataset from https://github.com/zalandoresearch/fashion-mnist/tree/master/data/fashion and add it in a folder data/fashion-mnist
  • Run test_fashion_mnist.m to generate the figures.
• Dispersion_Experiment: contains the code to generate figure 5 in the paper. In order to generate the figure, you need to go to the folder and then run:

julia generate_variances.jl
python darw_python.py

• HK_Power_Results: contains the code to generate the result reported in table 4.