Spatial Clustering Technique for Data Mining

Abstract

For mining features from the social web, analysis of the shape, detection of network topology and corresponding special meanings and also clustering of data become tools, because the information obtained by these tools can create useful data behind the social web by revealing its relationships and the relative positions of data. For example, if we want to understand the effect of someone’s statement on others, it is necessary to analyze the total interaction between all data elements and evaluate the focused data that results from the interactions. Otherwise, the precise effect of the data cannot be obtained. Thus, the effect becomes a special feature of the organized data, which is represented by a suitable form in which interaction works well. The feature, which is included by social web and it is effect someone’s statement, may be the shape of a network or the particular location of data or a cluster. So far, most conventional representations of the data structure of the social web use networks, because all objects are typically described by the relations of pairs of objects. The weak aspect of network representation is the scalability problem when we deal with huge numbers of objects on the Web. It is becoming standard to analyze or mine data from networks in the social web with hundreds of millions of items. Complex network analysis mainly focuses on the shape or clustering coefficients of the whole network, and the aspects and attributes of the network are also studied using semistructured data-mining techniques. These methods use the whole network and data directly, but they have high computational costs for scanning all objects in the network. For that reason, the network node relocation problem is important for solving these social-web data-mining problems. If we can relocate objects in the network into a new space in which it is easier to understand some aspects or attributes, we can more easily show or extract the features of shapes or clusters in that space, and network visualization becomes a space-relocation problem. Nonmetric multidimensional scaling (MDS) is a well-known technique for solving new-space relocation problems of networks. Kruskal (1964) showed how to relocate an object into n-dimensional space using interobject similarity or dissimilarity. Komazawa & Hayashi (1982) solved Kruskal’s MDS as an eigenvalue problem, which is called quantification method IV (Q-IV). However, these techniques have limitations for cluster objects because the stress, which is the attraction or repulsive force between two objects, is expressed by a linear formula. Thus, these methods can relocate exact positions of objects into a space but it is difficult to translate clusters into that space. This chapter introduces a novel technique called Associated Keyword Space (ASKS) for the space-relocation problem,which can create clusters from object correlations. ASKS is based on 16