Concept Drift Detection based on Anomaly Analysis

Publication Type:
Conference
Citation:
2014
Issue Date:
2014-12-01
Full metadata record
Abstract. In online machine learning, the ability to adapt to new concept quick-ly is highly desired. In this paper, we propose a novel concept drift detection method, which is called Anomaly Analysis Drift Detection (AADD), to im-prove the performance of machine learning algorithms under non-stationary en-vironment. The proposed AADD method is based on an anomaly analysis of learner’s accuracy associate with the similarity between learners’ training do-main and test data. This method first identifies whether there are conflicts be-tween current concept and new coming data. Then the learner will incremental-ly learn the non-conflict data, which will not decrease the accuracy of the learn-er on previous trained data, for concept extension. Otherwise, a new learner will be created based on the new data. Experiments illustrate that this AADD meth-od can detect new concept quickly and learn extensional drift incrementally.
Please use this identifier to cite or link to this item: