The time complexity of most clustering algorithms depends on the dimensionality of the input data and thus most clustering algorithms are slow on
highdimensional data. To solve this problem, we trained a deep autoencoder
and used it to compress the input data into a lower dimensional space with information loss. We reimplemented and extended the DeepCluster framework
proposed by Tian et al [26]. The original framework supports only K-means
and GMM clusterings. We extended it with hierarchical clustering, DBSCAN, and ensemble clustering. We evaluated the clusters and interpreted
the autoencoder with constructive induction. Both frameworks proved to
be unsuccessful in our experiments. However, we were able to interpret the
model and visualize its knowledge
|
