SCAFE - Simultaneous Clustering And Feature Extraction

This repository contains the code we used to test the model pipeline that we proposed in our CISBAT paper (September 2019).

Model Overview

Figure above shows the overview of the pipeline that we use in our paper. We first encode the time-series into three-channel images where the channels are Markov Transition, Gramian Angular Summation and Gramian Angular Difference Fields. Next, we train an undercomplete autoencoder to learn the manifold of the underlying data-generating distribution. Finally, we use ward-linking agglomerative clustering method to cluster the instances in the latent space, based on the learned manifold.

Centroids

We obtained centroids similar to those shown bellow:

Features

We used a modified version of the Grad-CAM algorithm to visualize the features that contributed the most to the clustering:

Running the code

Create an Anaconda environment by running:

conda env create -f reqs.yml

Activate the scafe environment by executing:

conda activate scafe.

Finally, install pyts by executing

pip install pyts==0.7.5

Note that the data should be stored in a .csv file. To run the autoencoder training and clustering with feature extraction (full pipeline) run:

python scafe.py --full --path ./path-to-your-data-csv/.

We strongly recommend to train the model only on machines that have CUDA support.