The Study Of Selective Adaptive Ensemble Learning Method For Concept Drift Problem

In the age of big data,incremental learning as a method that can process data in-crementally is becoming more and more important,and concept drift is one of the key challenges needing to be solved in incremetal learning.So far,a lot of algorithms have been proposed to cope with it,but it is still difficult to respond quickly to the change of concept.In this study,we aim at solving the concept drift problem in incremental learning.First,a novel method named Selective Transfer Incremental Learning(STIL)is proposed to deal with this tough issue.STIL uses a selective transfer strategy based on the well-known chunk-based ensembel algorithm.In this way,STIL can adapt to the new concept drift of data well through transfer learning,and prevent negative transfer and overfitting that may occur in the transfer learning effectively by an appropriate selective policy.The algorithm was evaluated on 15 synthetic datasets and three real-world datasets,the experiment results show that STIL performs better in almost all of the datasets compared with five other state-of-the-art methods.Afer that,we study the effect of base model's adaptive to ensemble model.And through a series of experiment and anlysis we find that have a set of very strong adap-tive base model will restrict ensemble model's responding to concept drift instead of improving the performance of ensemble model.Based on this rule,we further find that compared to improve all the base model's adaptive,only improve parts of the base model's adaptive will strength the performance of algorithm more effectively.And this is more useful to the ensemble model that has high adpative base mode]afer improved.In the current study background that chunk-based ensemble model trends to adopt base models that could update incrementally,our study shows that the improvement of base models should be moderate and the parts' improvement could lead to a better result.What's more,our two perspectives is verified by 6 synthetic datasets and two real-world datasets,and the experiment results proved our finding very well.