ralyhu / cmab-cc

Code and data for the paper "A Combinatorial Multi-Armed Bandit Approach to Correlation Clustering", DAMI 2023

A Combinatorial Multi-Armed Bandit Approach to Correlation Clustering

Overview

This project is developed as part of the following research paper:

F. Gullo, D. Mandaglio, A. Tagarelli (2023). A Combinatorial Multi-Armed Bandit Approach to Correlation Clustering published in Data Mining and Knowledge Discovery (DAMI), 2023

Please cite the above paper in any research publication you may produce using this code or data/analysis derived from it.

Folders

• datasets: contains the original data as well as the preprocessed data (as described in the paper). Biggest networks (to be unzipped in datasets folder) can be downloaded at the following link
• code: it contains this project code
• output: it stores all results produced by the algorithms for each round/run, e.g. "avg_expected_lossses.txt" contains a row for each indipendent cmab run and each row contains, for each t=0,...,T-1 (separated with ";"), the expected cumulative (up to round t) average disagreement of the yielded clusterings.

Usage

From the folder 'CMAB-CC/code', run the following command:

run_CMAB.py [-h] -d DATASET_NAME [-b {cc-clcb,cc-clcb-m,global-clcb,global-clcb-m,eg,eg-fixed,pe,cts,pcexp-clcb}] [-eps EXPLORATION_PROBABILITY] [-o {pivot,charikar}] [-T TIMESTEPS]
                   [-r RUNS] [-s SEED]                            

Dependencies

• networkx==2.6.3
• numpy==1.22.1
• PuLP==2.6.0
• python_igraph==0.9.9
• scipy==1.7.3

Positional arguments

  -d DATASET, --dataset DATASET
                        Input dataset, whose name identifies a particular subfolder in 'datasets/'

Optional arguments

  -h, --help            show this help message and exit
  -b {cc-clcb,cc-clcb-m,global-clcb,global-clcb-m,eg,eg-fixed,pe,cts,pcexp-clcb}, --bandit {cc-clcb,cc-clcb-m,global-clcb,global-clcb-m,eg,eg-fixed,pe,cts,pcexp-clcb}
                        Bandit algorithm (default value cc-clcb)
  -eps EXPLORATION_PROBABILITY, --exploration_probability EXPLORATION_PROBABILITY
                        exploration probability for epsilon-greedy CMAB algorithm (default value 0.1)
  -o {pivot,charikar}, --oracle {pivot,charikar}
                        Oracle to use in each CMAB step (default value pivot)
  -T TIMESTEPS, --timesteps TIMESTEPS
                        Number of timesteps/rounds to run the selected bandit algorithm (default value 400)
  -r RUNS, --runs RUNS  Number of (independent) bandit runs (default value 5)
  -s SEED, --seed SEED  Random generation seed -- for reproducibility (default value 100)  

  Modify the file "constants.py" to change the default values for the parameters.