This readme describes the code for the FSR algorithm to mine significant patterns with few-shots resampling.

The FSR algorithm to mine significant subgroups can be executed with the script fsr-alg.py. It accepts several parameters that are listed and described with the -h option (see below for more details):

usage: fsr-alg.py [-h] [-db DB] [-target TARGET] [-tval TVAL] [-k K]
                  [-maxd MAXD] [-eps EPS] [-corr CORR] [-mine MINE] [-res RES]
                  [-p P] [-pn PN] [-dfs DFS] [-simp SIMP] [-wy WY] [-ub UB]
                  [-cond COND] [-o O] [-ores ORES] [-d D] [-cat CAT]
                  [-geneexp GENEEXP] [-head HEAD]

optional arguments:
  -h, --help        show this help message and exit
  -db DB            path to input file
  -target TARGET    string of target column
  -tval TVAL        value of target to consider
  -k K              number of top-k results to find (def.=10000)
  -maxd MAXD        max number of conjunction terms (def.=3)
  -eps EPS          default value of eps to use for output (ignored when
                    corr=1)
  -corr CORR        run correction (def.=1)
  -mine MINE        run mining (def.=1)
  -res RES          number of resamples (def.=10)
  -p P              parallel computations (def.=1)
  -pn PN            number of parallel cores (def.=0 use all)
  -dfs DFS          use dfs exploration (def.=1)
  -simp SIMP        use simple exploration (def.=0)
  -wy WY            run WY correction (def.=0)
  -ub UB            use union bound (def.=0)
  -cond COND        use conditional distribution correction (def.=0)
  -o O              output path
  -ores ORES        output path for significant subgroups (def. no output)
  -d D              delta (def. 0.05)
  -cat CAT          categorize data (def. 0)
  -geneexp GENEEXP  custom search space for gene expression data (def. 0)
  -head HEAD        1 if data has an header (def. 0)

The fsr-alg.py script accepts an input dataset as a comma separated value file. The path for the input file must be given with the -db argument. It is then necessary to provide the column name (or number) of the target feature (argument -target), and the target value to consider as value 1 (argument -tval), while all other values are set to target 0. The script assumes that the input dataset does not have an header to parse the column name. In case an header is included in the dataset, use the argument -head 1.

The fsr-alg.py script mines the k most significant subgroups, where k is specified with the argument -k (default set to 10000). The maximum number of conditions to include in any subgroup is specified with the argument -maxd (default set to 3). The number of resamples of the target labels is set with the argument -res (default set to 10). The script outputs the experiment statistics (running time, number of significant patterns, ...) in a comma separated .csv file with the path given to the argument -o (default set to results_signfsr.csv). The significant patterns can be saved to a .csv file using the argument -ores.

1. Download the datasets.zip archive from this link: https://tinyurl.com/FSRdatasets
2. move and unzip the archive in the /data/ folder
3. install the required python 3 packages with the command pip install -r requirements.txt
4. run all experiments of "Impact of parameters on FSR" paragraph with the command python run_experiments_params.py
5. run all experiments of "Evaluation of FSR-C" and "Evaluation of FSR-U" paragraphs with the command python run_all_experiments.py
6. run all experiments of "Application to Neural Network interpretation" paragraph with the command python run_mnist.py

You can contact us at leonardo.pellegrina@unipd.it and fabio.vandin@unipd.it for any questions and for reporting bugs.