Clustering Non-Stationary Data Streams with Online Deep Learning

Abstract: With more devices connected, sensor data logged and people active in social networks, the trend towards working with dynamic data is clear. The number of applications where it becomes essential to perform real time analysis on data streams grows accordingly, each with its own challenges. From this area of data stream analysis we benchmark the performance of current state of the art clustering algorithms: CluStream, DenStream and ClusTree. We also adapt a Variational Autoencoder to perform in the context of non-stationary data streams and assess its generative capabilities for dimensionality reduction. From this limited lab experiment we show that while there is a significant improvement in the clustering accuracy of high dimensional datasets after a dimensionality reduction with a Variational Autoencoder, not all clustering algorithms benefit in the same way from it. Additionally we show that regardless of the clustering algorithm, no relevant improvement in the purity of the clusters could be obtained after the dimensionality reduction.