Article Preview
Top1. Introduction
In the literature, clustering is the most commonly used data analysis technique. Standard clustering methods (such as K-Means, hierarchical clustering and self organizing maps) are partitions the set of objects into distinct groups called clusters based on their similarities existing across the features. Thereby, they may fail to uncover clusters that are similar only over some but not all features. In contrast, biclustering aims at finding subsets of users which are behaving similarly over a subset of pages. The usefulness of this concept in the context of microarray measurements has been demonstrated in different studies (Busygin et al., 2002; Cheng et al., 2000).
In the context of microarray analysis, biclustering was firstly considered by Cheng and Church in 2000. Cheng and Church (CC) algorithm (Cheng et al., 2000). is a greedy iterative search method and consists in building a bicluster adding or removing rows or columns iteratively, thus, improving its quality which is measured with the Mean Squared Residue (MSR). The MSR is based on the sum of the squared residues which measure how adequate each expression value is, in comparison with the rest of the values of the bicluster. In Getz et al. (2000), an iterative hierarchical clustering is separately applied to each dimension and biclusters are built by means of the combination of the obtained results for each dimension.
There is a group of biclustering algorithms based on meta-heuristics such as evolutionary approaches (Bleuler et al., 2004; Divina et al., 2006), multiobjective evolutionary approaches (Banka et al., 2006; Divina et al., 2006), Simulated Annealing (Bryan et al., 2005), Particle Swarm Optimization, greedy randomized adaptive search (Dharan et al., 2009), Estimation of Distribution Algorithms (Liu et al., 2006) or Memetics Algorithms (Gallo et al., 2009). All these algorithms used the MSR as a part of their fitness function. The MSR is effective in recognizing biclusters with shifting patterns but not some patterns with scaling trends, in spite of representing quality patterns. Aguilar-Ruiz et al. (2009) proved that the MSR is not a good measure in order to discover patterns in data when the variance of values is high, that is, when the users present scaling patterns. And a very few biclustering work has been done in the field of web usage mining. Xu et al. (2010), presented a co-clustering algorithm using bipartite spectral clustering to extract bicluster from web users and pages and the impact of using various clustering algorithms is also investigated in that paper. A novel web co-clustering algorithm named Co-Clustering in Semantic space (COCS) was proposed by Zong et al. (2010), which simultaneously partition web users and pages via a latent semantic analysis approach.