Honorius Gâlmeanu

In dynamic environments, predictive models face the challenge of concept drift - changes in data distributions that degrade model performance over time. Concept drift detection and visualization can bring valuable insight into the data dynamics, especially for multidimensional data, and are related to visual knowledge discovery. To address this, we introduce a novel visualization model based on parallel coordinates, Parallel Histograms through Time (PHT). It allows visualization and identification of concept drift in the input data, enabling the simultaneous observation of the evolution of several features. When paired with a classifier, the PHT model provides a simultaneous visualization of the data changes in terms of (i) histograms over time and (ii) variation of the characteristics of the features. Although a classifier can identify changes, the PHT model enhances this capability by visually representing the detected drift, offering good premises for a comprehensive justification of the observed changes. Experimental evaluations on synthetic (CIRCLES, SINE1) and real-world (WEATHER, ELECTRICITY) datasets demonstrate that our method detects drift with high accuracy. In conjunction with domain-specific causality analysis, our model can be used to justify the identified and visualized concept drifts.

Data classification in streams where the underlying distribution changes over time is known to be difficult. This problem—known as concept drift detection—involves two aspects: (i) detecting the concept drift and (ii) adapting the classifier. Online training only considers the most recent samples; they form the so-called shifting window. Dynamic adaptation to concept drift is performed by varying the width of the window. Defining an online Support Vector Machine (SVM) classifier able to cope with concept drift by dynamically changing the window size and avoiding retraining from scratch is currently an open problem. We introduce the Adaptive Incremental–Decremental SVM (AIDSVM), a model that adjusts the shifting window width using the Hoeffding statistical test. We evaluate AIDSVM performance on both synthetic and real-world drift datasets. Experiments show a significant accuracy improvement when encountering concept drift, compared with similar drift detection models defined in the literature. The AIDSVM is efficient, since it is not retrained from scratch after the shifting window slides.

Training a support vector machine (SVM) for regression (function approximation) in an incremental/decremental way consists essentially in migrating the input vectors in and out of the support vector set with specific modification of the associated thresholds. We introduce with full details such a method, which allows for defining the exact increments or decrements associated with the thresholds before vector migrations take place. Two delicate issues are especially addressed: the variation of the regularization parameter (for tuning the model performance) and the extreme situations where the support vector set becomes empty. We experimentally compare our method with several regression methods: the multilayer perceptron, two standard SVM implementations, and two models based on adaptive resonance theory.