رسائل دكتوراة
English
2022
Data Stream Mining Using an Evolving Intelligent System
حسين عبد الأمير عباس فهد, نبيل هاشم كاغد جارالله
كلية تكنولوجيا المعلومات-جامعة بابل · العراق
Most of real world applications are able to generate the data stream. Data streams have unique characteristics anddifferent issues as well as it changes over time.The evolving system is presented as a convenient solution to overcome the issues of stream miningdue to theability to change its structureaccording to change of data behavior.Therefore, this dissertation presents an evolving intelligent system that based on the clustering technique to solve some problems of the stream mining.In the proposed system, the evolving Cauchy (e-Cauchy) algorithm is applied. Despite this algorithm isa successful algorithm for data stream clustering, but suffers from the large number of generated clusters therefore, this algorithm is developed to overcome this limitation. Moreover, the general clustering problem is the evaluation, which is solved by using deep learning method, that means it is used as an external evaluator. However, this method increases the accuracy of the network and ensures its stability by training the network with a non-constant (dynamic) learning rate value.Accordingly, the proposed system consists of two models, deep learning model (off line mode) and evolving clustering model (online mode).This proposed system aims to make a decision based fuzzy rules that changes over time to assign the data stream sample to a cluster accurately. In addition, the proposed system is capable to
detect new data samples (that have not been previously trained) and then, create a new cluster.The number of stream datasets used to test the proposed system is eleven datasets that generated from different applications such as,sensor, monitoring systems, devices and diverse domains like, questionnaire, behavior recognition and human activity recognition.Moreover, these datasets have different environments, complexity, number of features and classes.The evaluation measurements for this system are, accuracy, precision, recall, F1-score and the silhouette coefficient.The accuracy, precision, recall and F1_score ranged from (88.75%, 88.74%, 88.75%, 88.74%)to (100.0%, 100.0%, 100.0%, 100.0%). While the improvement rate for the silhouette coefficient from 0.28 to0.99 among all datasets