CLUSTERING MOBILITY PATTERNS IN WIRELESS NETWORKS WITH A SPATIOTEMPORAL SIMILARITY MEASURE

Clustering is a technique in data mining whose task is to classify objects into groups. In the recent years, it has been utilized to predict mobility behaviors of users for improving the quality and the management of services in wireless networks. Most of the current solutions focus on extending the traditional k-means approach with the numerical data to the categorical ones. However, such an extension paradigm may result in the loss of semantics of the spatio-temporal mobility patterns of users in the wireless network. Moreover, applying the random choice of initial values (or seeds) of the k-means technique may produce a different local optimum in every run time and thus lead to various partitionings. In this paper, we first propose a model for estimating the similarity among mobility patterns based on the weighted combination of Spatial and Temporal Pattern Similarity measures (STPS) of mobile users in wireless networks. Then we introduce the algorithm of Similarity Mobility Pattern based Clustering (SMPC), which is an alternative extension of the traditional k-means technique. Our approach focuses on using the proposed measure STPS to define a new concept of "cluster center" and to construct two novel procedures: a center updating procedure and a seed initialization procedure. We have conducted experiments with various conditions and parameters to investigate the suitability of the proposed similarity measure STPS and the quality of clusters generated from the algorithm SMPC for mobility patterns in the wireless environment. Experimental results have demonstrated that: (i) Integrating the spatial and temporal characteristics of mobility patterns in the similarity model improves considerably the clustering quality; (ii) Our seed initialization and center updating procedures achieve the stability and the computational speed better than ones with the traditional random initialization; (iii) Our clustering algorithm SMPC with the proposed combination similarity measure is more effective in computation than the other ones.