Community Detection to Invariant Pattern Clustering in Images

Community detection is a kind of clustering task which aims to find groups of vertices densely connected internally but sparsely connected to other groups. In comparison with conventional clustering methods, community detection methods are able to examine structural, functional and dynamical properties of the networked data, beyond its physical attributes. However, such techniques have been barely explored in the literature as most of the machine learning data sets are represented as non-graph data (e.g., feature vectors, images, texts, and so on). In this work, we propose a simple community detection framework based on the literature that covers since the graph construction process from feature vectors generated from non-graph data until the application and evaluation of community detection methods over such a graph. The framework is further evaluated on the problem of invariant pattern clustering of images, which consists of given a set of image objects taken from different angles, positions or rotations, clustering the images related to each object. Experiments were conducted considering three community detection methods (fast greedy, walk-trap and label propagation) and two relevant clustering methods (k-means and HDBSCAN). The results indicate FG as the better choice among those algorithms as it usually approximates efficiently the ground-truth groups while also keeping a reasonable number of communities. Moreover, our results suggest that community detection may be an efficient task not only to cluster graph data, but also domain applications represented by non-graph data.