Concept Drift Detection via Improved Deep Belief Network
Nafiseh Hatamikhah, Morteza Barari, Mohammad Reza Kangavari, Mohammad Ali Keyvanrad · 2018
One of the issues raised in streaming data is concept drift detection. In fact, the process of concept drift comes from natural tendency events in the real world to change over time. For example, in data receiving from credit card transactions, detect when transactions rise suddenly, can help in identifying the fraud. In this paper regards to the importance of concept drift in streaming data, a solution to accurate diagnosis and timely is presented. This solution is based on ensemble algorithm and “streaming ensemble algorithm” (SEA) algorithm that SEA algorithm is used as one of the most commonly stream algorithms. This approach uses a deep belief network as a basic model in the SEA algorithm. In the method which is presented in this paper, we used the change of classification error on new data for concept drift detection. Analyzing the results shows that the proposed method compared with similar algorithms, in addition to a significant reduction in the runtime, improved F_measure criteria.