A study of feature representation via neural network feature extraction and weighted distance for clustering

Neural Networks are well known for its performance to classify and cluster data sets via multiple layers of networks passing and transforming information pictured by raw data. The feature layer projects the raw data into a space spanned by hidden features. To understand data representations in both original (i.e., image) and feature spaces, the main purpose of this research is to analyze the clustering performance with different feature representations. Naturally, distance measures have a great impact on clustering performance. Different distances and their combinations are tested on both the original and feature spaces. The combined distances were obtained by using different optimal weights that minimize classification errors in different measures via a series of optimization models. These weights were multiplied by their respective distances in order to create the combined distance. Clustering was evaluated using silhouette scores. The feature space in general has better performance, in terms of clustering, than the image space, with Cosine Similarity being the best distance for both the image space and feature space.