Joint Optimization of an Autoencoder for Clustering and Embedding -- Code

This is the code corresponding to the experiments conducted for the work "Unsupervised Scalable Representation Learning for Multivariate Time Series"

Requirements

Experiments were done with the following package versions for Python 3.8.3:

• Numpy (numpy) v1.18.1;
• Scikit-learn (sklearn) v0.22.1;
• Tensorflow (tensorflow) 2.2.0.

Note that dkm.py requires an earlier version of Tensorflow

• Tensorflow (tensorflow) 1.15.0.

Datasets

The datasets will be generated using python make_datsets.py

Usage

To train the clustering module (CM) on a dataset $DATASET with $INITIALIZATION={rand|pre} (random or pretrained) initialization:

python CM.py $DATASET $INITIALIZATION -b BATCHSIZE -e EPOCH -p PRETRAINING_EPOCH -r NB_RUNS -g GPU_ID

The same function can be used for AECM.py, DCN.py, DEC.py and IDEC.py

The results will be saved for example in $DATASET/save/save-cm-rand.npz.

That archive contains 12 files:

• llk value of the loss,
• wgt weights of the trained network,
• lbl predicted labels,
• klbl predicted labels by k-means on the embedding,
• ari,nmi,acc Adjusted Rand index, Normalized Mutual Information and Accuracy of the predicted label (lbl),
• kari,knmi,kacc Adjusted Rand index, Normalized Mutual Information and Accuracy of the predicted label by k-means (klbl),
• epc Duration of the training.

To train DKM, dkm.py accordingly to the indications provided in the [authors' repository].