Partition2Vec - Towards Representation Learning for Evolutionary Clustering using Siamese Autoencoders

We consider the problem of representation learning for evolutionary clustering. Evolutionary clustering algorithms are plagued by the difficulty of defining representations that are non-redundant, effective under crossover and avoid the introduction of strong assumptions regarding relevant cluster properties. At the core of our paper is a simple question: could recent advances in deep learning be leveraged to identify new representations for this unsupervised learning setting? We describe a Siamese Autoencoder as a first attempt to investigate this question on small toy problems and proceed to analyze the assumptions, generalization performance and limitations of the approach.