Macro SOStream: um algoritmo evolutivo para agrupamento autoorganizado baseado em densidade

Abstract

Situations that generate a continuous data stream under a changing environment, such as TCP / IP traffic, e-commerce, and industrial monitoring, can make the usability of offline algorithms for machine learning infeasible. That is due to the need for data storage, infinite growth of data generation, and memory restrictions, making models impossible to redesign and retrain. As a consequence of that, algorithms that work in a fully or partially on-line fashion have arisen. Among them, evolving algorithms is a prominent approach because such algorithms can develop and update models’ parameters and structure in unknown environments and detect concepts drift and evolution in the input-output data over time. Because evolving approaches and models are highly applicable in real-world problems, this work proposes a new evolving clustering algorithm named Macro SOStream. This algorithm performs on-line learning and is based on self-organizing density for data stream clustering. The Macro SOStream is based on the SOStream algorithm; however, we incorporate the notion of macroclusters as a microclusters’ composition. While microclusters have spherical shapes, macroclusters may assume arbitrary shapes. For the experiments, were used the benchmark datasets quite usual in the literature, the clustering performance metric Adjusted Rand Index (ARI), and is performed at the average time of the algorithms’ execution with the datasets. The results indicated that the performance and average execution time of the Macro SOStream algorithm are comparable to those of SOStream and DenStream.