guillaume-bosc / mcts4dm

Monte Carlo tree Search for Subgroup Discovery

MCTS4DM ReadMe

Introduction

MCTS4DM is a anytime pattern mining algorithm that is able to consider tabular data with numerical and categorical attributes and to extracting patterns that distingishe a target (subgroup discovery, exceptional model mining, ...).

Repository contents

To run MCTS4DM, you need to download the projet from this GitHub repository. Below, we detail the content of each folder and file contained in this repository.

• The "src" folder contains the source code of the MCTS4DM algorithm.
• The "datasets" folder contains the benchmark datasets used in our experiments. Each folder in the "dataBenchmark" folder is related to a dataset, and it contains three files: the files corresponding to the targets and the attributes in the CSV format, and the arff format of the dataset (to compare with other existing algorithms).
• The "results" folder contains the result of the runs.
• The "lib" folder contains the required libraries.
• The "MCTS4DM.jar" file that is the jar of the MCTS4DM algorithm.
• The "parameters.conf" file that is the configuration file for parameters that has to be passed as argument of the MCTS4DM algorithm.
• The "RunExperiments" folder contains the program that can launch automatically a batch of experiments. In this folder there are (i) the "results" folder containing the results of the batchs of experiments given with a defined hierarchy, (ii) the "src" folder containing the source codes of this program, (iii) the "RunExperiments.jar" file, and (iv) the "paramGen.conf" file generating during the program. To run this program, execute java -jar RunExperiment.jar [Options], where the options are chosen among {ucb, expand, rollout, memory, update, iter} that enables to run the experiment of a defined strategy of MCTS.
• The "ComparisonsWithCompetitors" folder contains the program that compare MCTS4DM with existing approaches, namely, SD-Map* (exhaustive search), beam search with the Cortanatool, SSDP (evolutionary algorithm) and Misere(sampling approach).
• The "GenerateDataPlot" folder contains the program to generate the data file from the results of MCTS4DM (located in RunExperiments/results with the specific hierarchy). For that, run the "GenerateDataPlot.jar" file by java -jar GenerateDataPlot.jar. To plot the graphes with GnuPlot, run the different script files ./launchQuality.sh, ./launchRuntime.sh and ./launchLength.sh located in the RunExperiments/results/ folder. The graphes will be created in each RunExperiments/results/<NameExperiment>/<Dataset>/ folder.

How to launch MCTS4DM

This part details the process to run MCTS4DM.

1. Format the dataset. A dataset is represented by two files : (i) the "properties.csv" file that contains the table of attributes where each line is an object and each column is an attribute; and (ii) the "qualities.csv" file that contains the table of labels where each line is an object (in the same order that in the "properties.csv") and each line is a target attribute. These file are in the CSV format where the separator character between columns is "\t".

2. Fill out the parameters configuration file. Create and fill out a parameters configuration file exactly respecting the format of the "parameters.conf" file.

3. Run MCTS4DM. Once the dataset files and the parameters configuration file are completed, run MCTS4DM by: java -jar MCTS4DM.jar <Path to the parameters configuration file>. For example: java -jar MCTS4DM.jar parameters.conf

4. Get the results. Once the execution is over, the results are available and in the resultFolderName (provided in the parameters configuration file) within the "results" folder. Four files are created: (i) the "info.log" that contains the values taken by the parameters, the runtime and the date of the run; (ii) the "results.log" file that where each line contains a subgroup. The format is : <Description>\t<Targets>\t<Measure>\t<E11>\t<E10>\t<E01> where E11 is the number of objects that respect the description and that are associated to the targets, E10 is the number of objects that respect the description and that are not associated to the targets, E01 is the number of objects that do not respect the description and that are associated to the targets. (iii) the "support.log" file that contains the support of each subgroup (the ith line correspond to the support of the ith subgroups of the "result.log" file), and (iv) the "supportE11.log" file that only contains the IDs of the objects in E11.

For any questions/remarks, contact Guillaume BOSC: [email protected] or Mehdi Kaytoue: [email protected]