One of the most important challenges for\nmachine learning community is to develop efficient classifiers\nwhich are able to cope with data streams, especially\nwith the presence of the so-called concept drift. This phenomenon\nis responsible for the change of classification task\ncharacteristics, and poses a challenge for the learning model\nto adapt itself to the current state of the environment. So there\nis a strong belief that one-class classification is a promising\nresearch direction for data stream analysisââ?¬â?it can be used for\nbinary classification without an access to counterexamples,\ndecomposing a multi-class data stream, outlier detection or\nnovel class recognition. This paper reports a novel modification\nof weighted one-class support vector machine, adapted\nto the non-stationary streaming data analysis. Our proposition\ncan deal with the gradual concept drift, as the introduced\none-class classifier model can adapt its decision boundary to\nnew, incoming data and additionally employs a forgetting\nmechanism which boosts the ability of the classifier to follow\nthe model changes. In this work, we propose several\ndifferent strategies for incremental learning and forgetting,\nand additionally we evaluate them on the basis of several real\ndata streams. Obtained results confirmed the usability of proposed\nclassifier to the problem of data stream classification\nwith the presence of concept drift. Additionally, implemented forgetting mechanism assures the limited memory consumption,\nbecause only quite new and valuable examples should\nbe memorized.
Loading....