Tensorflow implementation of the Dissimilarity Mixture Autoencoder:

• Juan S. Lara and Fabio A. González. "Dissimilarity Mixture Autoencoder for Deep Clustering" arXiv preprint arXiv:2006.08177 (2020).

In this paper, we introduce the Dissimilarity Mixture Autoencoder (DMAE), a novel neural network model that uses a dissimilarity function to generalize a family of density estimation and clustering methods. It is formulated in such a way that it internally estimates the parameters of a probability distribution through gradient-based optimization. Also, the proposed model can leverage from deep representation learning due to its straightforward incorporation into deep learning architectures, because, it consists of an encoder-decoder network that computes a probabilistic representation. Experimental evaluation was performed on image and text clustering benchmark datasets showing that the method is competitive in terms of unsupervised classification accuracy and normalized mutual information.

If you have anaconda installed, you can create the same environment that we used for the experiments using the following command:

conda env create -f dmae_env.yml

Then, you must activate the environment:

source activate dmae

or

conda activate dmae

This implementation is based on tf.keras.layers, therefore, DMAE can be easily used in other deep learning models as an intermediate layer. A replication of the experiments can be found in the folder examples, we highly recommend to check the experiments on synthetic data first: examples/synthetic/. They provide an interactive experience that is useful to interpret the learned representations of DMAE.

For the real data examples/real, you can run python experiments.py -h for more information about the possible options.

Some examples are:

• Quick test:

python experiments.py --trials 1 --pretrain_epochs 1 --cluster_epochs 1

• Replication of the MNIST results using the euclidean dissimilarity:

python experiments.py --dataset mnist --trials 10 --pretrain_epochs 500 --cluster_epochs 300 --da True --train_batch 256 --test_batch 1000 --dis euclidean

• Replication of the MNIST results using the mahalanobis dissimilarity:

python experiments.py --dataset mnist --trials 10 --pretrain_epochs 500 --cluster_epochs 300 --da True --train_batch 256 --test_batch 1000 --dis mahalanobis

• Replication of the Reuters experiments:

python experiments.py --dataset reuters10 --trials 10 --pretrain_epochs 100 --cluster_epochs 100 --train_batch 256 --test_batch 1000 --dis euclidean