Learning Predictive Models with a Triclustering-based Classifier

This work proposes a new class of explainable predictive models from three-way data using discriminative triclusters. Triclustering searches are dynamically hyperparameterized to comprehensively find for informative triclusters (groups of individuals with a coherent pattern along a subset of variables and time points) and then use these temporal patterns as features within a state-of-the-art classifier with strict guarantees of interpretability. To this end, a temporally constrained triclustering algorithm, termed TCtriCluster algorithm, is devised to mine time-contiguous triclusters.

Getting Started

These instructions will get you how to run the triclustering-based classifier. See the demo section for notes on how to run the demo project on your machine.

Prerequisites

To run the triclustering-based classifier you need to have Python 3.4 or above installed as well as the following packages:

• scikit-learn
• scipy
• numpy
• pandas
• sortedcontainers

How To Run

First, you should perform triclustering in your data.

Run the following command to see the arguments list for TCtriCluster.

python3 TCtriCluster.py -h

Run the triclustering algorithm with the your defined input parameters and save the result for an output file (.txt):

python3 TCtriCluster.py -f <input_file> -sT <min_t> -sS <min_s> -sG <min_g> -w <win_ratio> -o <opc> [-mv <mv_threshold>] > <output_file>.txt

Next, with the produced triclusters, you compute the similarity matrices:

python3 compute_similar_mats_tri.py <datafile> <target_column> <triclusters_file> <matrix_output> <categorical_features> <continuos_features>

Finally, the classifier uses the similarity matrices as the learning examples. This code performs n x k-fold Stratified CV to evaluate the performance of the classifier.

python3 compute_predictions.py <matrix_input_folder> <output.csv> <k-splits> <n_repeats>

Demo Example

We provide a demo example in demo folder.

First we parsed the d1.csv file to achieve the required formatting by the triclustering algorithm. To do this, use the following command:

# Usage python3 src/write_tab_file.py <csv_file> <output_tabfile> <target_var> <n_timepoints>

$ python3 src/write_tab_file.py demo/d1.csv demo/tab_file.tab Class 3

Next we performed triclustering with TCtriCluster algorithm:

$ python3 src/TCtriCluster.py -f demo/tab_file.tab -sT 2 -sS 2 -sG 2 -w 0.1 -o 1 > demo/triclusters_d1.txt

Next, with triclusters and the original dataset we computed the similiarity matrices

$ python3 src/compute_similar_mats_tri.py demo/d1.csv demo/triclusters_d1.txt demo/sim_matrices Class 3 [S1, S2, S3, S4, S5, S6] []

Finnally with the matrices we run the classifier, evaluating the results with repeated stratified k-fold CV:

$ python3 src/compute_predictions.py demo/sim_matrices Class 3 results.csv 2 2

Authors

• Diogo F. Soares
• Advised by Sara C. Madeira