A Comprehensive Review on Evolving Data Stream Clustering

Most of the real-world applications deal with large data streams that are generated continuously. With this increase in the size of data streams, the existing data stream clustering algorithms cannot be implemented efficiently as these are designed to tackle static data. This invalidation of the existing algorithms becomes even more prominent when dealing with continuous as well as evolving data streams. To effectively cluster such data streams, there is a need for algorithms that can deal with the memory and computational limitations along with the concept drift. In this regard, various state-of-the-art clustering techniques have been modified to tackle the continuity and evolving nature of data. One of the main approaches used by researchers is devising a hybrid framework, commonly termed as online-offline phase clustering. This approach effectively addresses Big Data. However, it does not provide any mechanism to detect the expired data points – data that has lost its significance due to concept drift. To further improve clustering algorithms, researchers came up with window models that serve as updating units, thus, successfully handling concept drift. Realizing the need for enhanced algorithms, in this paper we closely analyze these clustering algorithms by providing a detailed discussion on their performance from different perspectives, such as memory and computational cost, cluster quality and purity, scalability, effectiveness in detecting outliers, and the cluster shape. Our survey encompasses a comprehensive and thorough literature review of the evolving data stream clustering techniques established in recent years. Further, the paper highlights the shortcomings of the algorithms, and tends to target the limitations that should be tackled in future.