基于图的概念重现发现与预测

Recurring Concept Detection and Prediction Based on the Graph

概念漂移是数据流挖掘中具有挑战性的问题.当概念漂移发生后,原有分类模型的分类正确率会显著下降,因此需要及时发现并调整模型以适应这些改变.概念重现是概念漂移的特殊情况,然而已有的算法大多未能充分考虑这种状况.为此,提出一种能够处理重现的概念检测方法.试验结果表明,该方法能够以较低的延迟和较低的误报率检测到概念漂移,并且可以识别重现的概念,很大程度上提升了分类器的分类正确率.

Concept drift was a challenging problem in stream mining. When the concept drift occured, the ac- curacy of the original predictive model may decrease significantly. So it was necessary to put forward a feasible method to detect concept drift. Recurring concept is a special ease of concept drift. However, most of existing algorithms have not taken full account of this case. This research proposed an approach to the recurring con- cept detection problem. Extensive experiment revealed that the method we proposed could detect not only the concept drift with relatively low delay and rate of false positive, but also the recurring concepts. Moreover, the accuracy of the classification would be greatly improved.