Context aware real-time clustering with cortical coding method

Abstract

This thesis introduces a novel online clustering algorithm based on the cortical coding method to address the challenges of data stream clustering. Traditional clustering methods often struggle to deliver optimal performance in such scenarios. The proposed algorithm processes incoming data sequentially, creating new clusters or adapting existing ones while aiming to maximize the entropy. It incorporates an energy-based technique to regulate the creation, evolution, and removal of clusters. Additionally, the energy-based evolution mechanism for clusters removes the harmful effects of anomalous points. The motivation behind this research stems from the increasing demand for real-time systems that deal with large amounts of data. Traditional clustering methods lack scalability and require prior information, such as the number of clusters or a similarity threshold, for optimal performance. However, the optimal number of clusters and their characteristics can change over time in dynamic systems. Therefore, an ideal clustering method for dynamic data should be online, adaptive, and resilient to outliers. This thesis aims to present an extended version of the cortical coding method as a novel online clustering algorithm. The research evaluates the cortical coding method's clustering performance and time complexity on various problems, demonstrating its superiority over popular online clustering methods and comparable or superior performance to conventional offline methods. The proposed method shows promising results for the potential applications of this method in real-time anomaly detection, feature extraction, and state detection. Additionally, the method is evaluated on a real world anomaly detection problem to demonstrate its applications. In conclusion, the research presents an extended version of the cortical coding method as an effective solution for online clustering problems. The method demonstrates superior performance to compared methods and outperforms existing online clustering algorithms. It exhibits robustness against anomalous data and provides dynamic adaptation to changing systems. The article suggests potential future work, including further parameter analysis, exploration of parallelization techniques, and the method's application in anomaly detection and feature extraction.