A novel clustering method suitable for clustering of biological signal datasets containing batched outliers

Abstract

During clustering analyses, instances of batched outliers of one class falling close to another class can be a significant problem. Such outliers might be incorporated into a false class or lead to the false identification of unreal classes, which can lead to false localization of the cluster centers. Here we propose a novel method for accurate classification of outliers in batched clustering analyses, aimed specifically at the type of outliers most often encountered in biological signals. The recommended divisive hierarchical clustering method is based on how much each element in the dataset is unwanted by other elements. In this method, the reluctance vectors applied to each element by the other elements are first determined. According to the common features of the reluctance vectors (horizontal and vertical components), two initial classes are obtained from some elements. All remaining elements are then included into classes according to their proximity to these classes. Then, using the reluctance vectors developed between the two established classes, class that might be re-divided are identified and further classes are constituted using the same splitting method. To validate this approach, which we named the selfish data clustering (SDC) method, areal dataset was analyzed using the SDC method and other commonly applied clustering methods. We found that our clustering method outperformed the conventional approaches by up to 12% (average is 6%) in datasets with low silhouette values.