An Incremental Linear Discriminant Analysis for Data Streams Under Non-stationary Environments
Annie anak Joseph, Young-Min Jang, Seiichi Ozawa, Minho Lee
In real life, data are not always generated under stationary environments. However, traditional learning systems have normally assumed that the property of data streams is stationary over time, and this sometimes leads to the degradation in the system performance when there are some hidden contexts changes (e.g. changes in class boundaries and temporal trends in time series). Such context changes are called concept drifts, and various methods to handle concept drifts have been developed in machine learning and data mining fields. However, most of them are aiming for building classifier models. Considering that class boundaries have changed over time under non-stationary environments, extracted features should also be adapted to concept drifts autonomously. In this paper, we propose an extension of incremental linear discriminant analysis (ILDA) as an online feature extraction method under non-stationary environments. The extended ILDA has the following two functions: concept-drift detection and knowledge transfer. The recognition performance of the extended ILDA is evaluated for three benchmark data sets. Experimental results demonstrate that the recognition performance in the extended ILDA is greatly improved by introducing the knowledge transfer after the concept-drift detection.