E-STRSAGA: an ensemble learning method to handle concept drift

We present E-STRSAGA, an ensemble learning algorithm, that can efficiently maintain a model over a stream of data points and recover from any type of drift that may happen in the underlying distribution. This algorithm adopts the new distribution by efficiently adding new experts after detecting any change in the performance of its model, and forgets about the previous distribution by efficient way of dropping old experts and data points from the old distribution. Experimental results are provided on a variety of drift rates and types (abrupt, gradual and multiple abrupt drifts). Results confirm the competitiveness of E-STRSAGA with a streaming data algorithm that knows when exactly drift happens and is able to restart its model and train it only over new distribution