Publication

Abstract : Partitioning a set of objects in a particular domain into different subsets or clusters can be done using unsupervised learning methods. The algorithms compete to increase the intra-cluster similarity with an objective to find top-quality clusters in a limited time frame. High-dimensional data are very relevant to a wide range of areas but to make analysis and easy understanding we need to converge it to a low-dimensional representation. Here, we implement a spectral clustering algorithm for document data set in which we are able to cluster semantically similar documents with respect to the presence of semantically similar terms. Some documents will be similar in a particular subspace but dissimilar in other. The documents that belong to a union of low-dimensional subspaces are found from a collection. We get a cluster map which has a weighted link between the nodes indicating the strength of similarity between clusters in different subspaces. We apply principal component analysis (PCA) as a feature extraction step. There are no assumptions made about the structure or shape of the clusters. We analyse the cluster formation using homogeneity, inertia and Silhouette score for varying parameters (Epsilon-neighbourhood graph/K-nearest neighbour/fully connected, epsilon value, number of clusters) of spectral clustering.