Life-long learning based on dynamic combination model

In this paper, we propose a novel life-long learning framework, which constantly evolves with changing data distribution, learning new knowledge while retaining some old knowledge. In many practical systems, data in the past is still useful but no longer available. Therefore, a question arises on how to update the model based on both new data and current model. To address this issue, our framework lays its basis on ensemble method with multiple sub-classifiers, independent of base type. When new data is processed, new sub-classifiers are generated accordingly. The classifiers are then dynamically combined using decision tree, together with a novelly proposed pruning method to prevent overfitting and eliminate out-dated models. Guarantees are provided to the combination method. Experiments indicate that the framework achieves good performance when the data changes with time, and has better accuracy compared to existing transfer and incremental learning, and methods in stream data mining. (C) 2017 Elsevier B.V. All rights reserved.