Using Clustering Method to Construct A Classifier on Concept Drifting Streams

[[abstract]]Differ from the static database for storing history data, the data stream is continuously and unlimited produced in high-speed. Moreover, the implicit concept and the distribution of data may change as time goes by. Accordingly, the classification model is not only required to perform the predictions correctly and efficiently, but also to detect concept changes for adjusting the classification rules to catch recent trends in time. In this thesis, a clustering based classification method is provided for reducing the number of classification rules. First, the nearest neighbor algorithm is adopted to cluster the training data. Then a representational pattern is chosen from each cluster to construct a classification rule. In the process of predicting, the exception data in a sliding window is recorded and the accuracy in the window is monitored. When the accuracy is below a user-defined threshold value, the recorded data is used to adjust the classifier. The proposed method controls the number of classification rules by setting different distance threshold when performing data clustering. Therefore, the required storage of the constructed classifier is adaptable. Furthermore, the constructed classifier model is adjustable for concept changes to keep accurate predictions.