Weblog Clustering in Multilinear Algebra Perspective (0909.2345v1)

Abstract: This paper describes a clustering method to group the most similar and important weblogs with their descriptive shared words by using a technique from multilinear algebra known as PARAFAC tensor decomposition. The proposed method first creates labeled-link network representation of the weblog datasets, where the nodes are the blogs and the labels are the shared words. Then, 3-way adjacency tensor is extracted from the network and the PARAFAC decomposition is applied to the tensor to get pairs of node lists and label lists with scores attached to each list as the indication of the degree of importance. The clustering is done by sorting the lists in decreasing order and taking the pairs of top ranked blogs and words. Thus, unlike standard co-clustering methods, this method not only groups the similar blogs with their descriptive words but also tends to produce clusters of important blogs and descriptive words.