Soft-Link Spectral Clustering for Information Extraction

Unsupervised spectral clustering methods can yield good performance when identifying crisp clusters with low complexity since the learning algorithm does not rely on finding the local minima of an objective function and rather use spectral properties of the graph. Nonetheless, the performance of such approaches are usually affected by their uncertain parameters. Using the underlying structure of a general spectral clustering method, in this paper a new soft-link spectral clustering algorithm is introduced to identify clusters based on fuzzy k-nearest neighbor approach. We construct a soft weight matrix of a graph by identifying the upper and lower boundaries of learning parameters of the similarity function, specifically the fuzzifier parameter (fuzziness) of the Fuzzy k-Nearest Neighbor algorithm. The algorithm allows perturbations on the graph Laplace during the learning stage by the changes on these learning parameters. With the empirical analysis using an artificial and a real textual entailment dataset, we demonstrate that our initial hypothesis of implementing soft links can improve the classification performance of final outcome.