A self-adaptive insert strategy for content-based multidimensional database storage

Abstract

In this paper, we present the current development progress of our dynamic insert strategy based on the Intelligent Cluster Index (ICIx), which is a new type of multidimensional database storage. Opposite to purely value-based interval methods, ICIx performs a semantic clustering of the data objects in a database and keeps the clustering results as basis for storing in a special tree structure (V-Tree). Our paper aims at the quality problem caused by a trade-off between the static clustering that results from the initial training data set and the continuous insertion of data into a database which requires a continuous classification. The strategy that we propose will solve this problem through a continuous and effcient content-based growing of the initially static clustering. We have developed an additional structure - the C-Tree - which stores the knowlege of the hierarchical clustering component, i.e. hierarchical Growing Neural Gas (GNG), for unsupervised content based classification. In contrast to other methods (e.g. dynamic versions of R-Trees) we use the C-Tree to process the new tuple. Furthermore, we use a Bayesian approach to determine the degree of adaptation of the knowledge base. Using this value, we update the knowlege base and propagate the resulting changes to the V-Tree. As a result, we obtain a continuous content-based growing.