To partition a set of samples in a dataset, any clustering algorithm considers all features of dataset equivalently during the clustering procedure. It is highly likely that some features have more information in any dataset. If any clustering algorithm lets different features have unequal contributions in the final partitioning, it can highly likely result in a better set of resultant partitions. But the problem of how the features can participate in the clustering task in a weighted manner is a very challenging problem. The other problem is the mechanism of the weight assignment. Recently the problem has been dealt with by a Locally Adaptive Clustering (LAC) algorithm. However, like other traditional clustering algorithms, LAC algorithm is inefficiency when the at-hand dataset is imbalanced. In this paper a novel method is proposed which deals with both raised problems (that have not solved by LAC algorithm) while it exploits LAC algorithm privileges. While LAC algorithm forces the sum of weights for the clusters to be equal, our method lets them be unequal. This makes our method more flexible to escape from falling at the local optima. It also lets the cluster centers to be efficiently located in the feature space at fitter places. It finds better cluster centers than the ones found by its rivals. The proposed clustering algorithm is called Weighted Clusters in Locally Adaptive Clustering (WLAC). The paper shows the effectiveness of WLAC algorithm both theoretically and experimentally.

Login Required

Please login to access the full PDF of this article.

🔑 Author Login