a novel memory management strategy for efficient detection of concept evolution over stream data

نوع: Type: Thesis

مقطع: Segment: masters

عنوان: Title: a novel memory management strategy for efficient detection of concept evolution over stream data

ارائه دهنده: Provider: Mohammad Yasin Asgari

اساتید راهنما: Supervisors: Morteza Yousef Sanati

اساتید مشاور: Advisory Professors: Moharam Mansoori Zadeh

اساتید ممتحن یا داور: Examining professors or referees: Mehdi Sakhaee Nia AND Reza Mohammadi

زمان و تاریخ ارائه: Time and date of presentation: 2025

مکان ارائه: Place of presentation: seminar

چکیده: Abstract: In recent years, emerging class detection in data streams has become a significant challenge in machine learning, particularly in dynamic and evolving environments. This thesis proposes a novel framework for accurate and efficient detection of emerging classes based on clustering, the Local Outlier Factor (LOF), and continuous base model updates. One of the key innovations of this study is the design and implementation of an intelligent strategy for selectively removing low-value instances from the buffer. This aims to improve system efficiency without compromising the accuracy of novel class identification. In the proposed approach, only instances with a LOF below a defined threshold are removed, as such instances are considered to belong to known classes and are unlikely to contribute to the discovery of emerging ones. The framework is evaluated on standard datasets including Shuttle, KDDCup99, and PAMAP. Experimental results demonstrate improved performance in terms of lower error rate and higher precision in novel class detection compared to previous methods. Additionally, it is shown that the selective removal strategy minimally affects evaluation metrics such as M_new, F_new, and Error. Finally, future work directions are discussed, including buffer partitioning and dynamic threshold tuning for further improvements.