A new approach to improve the accuracy of online clustering algorithm based on scatter/gather model

Abstract

In cluster analysis process used in data mining which enables extracting interesting data patterns from datasets, accuracy and efficiency are the factors which play a pivotal role. Scatter/Gather is a cluster-based browsing model, and most of previous works on this model focused on efficiency of the clustering algorithm. In this paper we present an algorithm which could improve the accuracy of the online clustering algorithm while still maintain a reasonable level of efficiency. Our experiment proves that the new algorithm is more accurate than the original algorithm.

A Cluster-Based Browsing Model For QoS-Aware Web Service Selection

“…From this perspective, NLAour improved version of LAIR2 algorithm, is the most efficient implementation of Scatter/Gather model for the QoS interval dataset. It reconfirms the conclusion from [13,14].…”

“…We have four different implementations of the browsing model based on two originals, and two improved algorithms. The original algorithms consist of the original Scatter/Gather and the iterative Symbolic Dynamic Clustering Algorithm (SCLUST) [12]; furthermore, our improved algorithms include the improved SCLUST and the Hk-means (NLA) algorithm [13], which is implemented based on the improvement on the LAIR2 algorithm [14]. For the improved techniques, we intended to increase the accuracy, as well as the efficiency (based on the processing time) of our clustering results.…”

“…For instance, as shown in Figure 5 (1), the desired number of clusters is chosen as three. Hence, in level 2 of the hierarchy, once the first two clusters are found, the system is unable to select which cluster should be split first, and to which sub-cluster it should move (to return the three desired clusters), in order to have the most accurate results [13].…”

“…Figure 5-How to Select Clusters along the Hierarchy [13] Hence, in contrast to the LAIR2 algorithm, in which the search moves to the next level of hierarchy after all clusters in previous levels are split, our proposed system is able to choose any sub-cluster in any levels, and moves down in a single branch as deep as necessary to find the cluster with the lowest homogeneity [13].…”

“…However, in our proposed approach, the search always begins from the minimum index of the clusters selected by the user (which was specified for each cluster in the previous phase) in the hierarchy. Thus, the duration of the search process would be almost the same at any iteration, even for a dataset with a large amount of data [13].…”