Concept Drift Visualization of SVM with Shifting Window

TL;DR

This work tackles the challenge of detecting and explaining concept drift in streaming data. It introduces Parallel Histograms through Time (PHT), a visualization framework that extends parallel histograms with time-unfolded windows to display per-feature distributions and their shifts, including per-class perspectives. A meta-algorithm is proposed to localize drift points, and the approach is demonstrated with an incremental/decremental SVM with shifting window on datasets such as Forest Covertype, SINE1, and CIRCLES, showing both drift detection and localization. The method aims to support visual knowledge discovery by providing interpretable drift explanations that can guide model adaptation and selection of windowing parameters.

Abstract

In machine learning, concept drift is an evolution of information that invalidates the current data model. It happens when the statistical properties of the input data change over time in unforeseen ways. Concept drift detection is crucial when dealing with dynamically changing data. Its visualization can bring valuable insight into the data dynamics, especially for multidimensional data, and is related to visual knowledge discovery. We propose a novel visualization model based on parallel coordinates, denoted as parallel histograms through time. Our model represents histograms of feature distributions for successive time-shifted windows. The drift is shown as variations of these histograms, obtained by connecting the means of the distribution for successive time windows. We show how these diagrams can be used to explain the decision made by the machine learning model in choosing the drift point. By isolating the drift at the edges of successive time windows, there will be none (or reduced) drift within the adjacent windows. We illustrate this concept on both synthetic and real datasets. In our experiments, we use an incremental/decremental SVM with shifting window, introduced by us in previous work. With our proposed technique, in addition to detect the presence of concept drift, we can also depict it. This information can be further used to explain the change. mental results, opening the possibility for further investigations.

Figures (5)

• Figure 1: The seven most representative features of the CIRCLES datasetIEEEexample:Bifet2006 are represented. Parallel histograms consist of a parallel coordinate system where each vertical axis also contains a representation of the probability density (more often just a histogram). The height of each bin is relative to the height of the highest bin of that axis, and thus the histograms are not comparable among themselves. Plots in (a) and (b) represent the same dataset, but different time windows. The per-feature distributions are different in (a) vs. (b), revealing a concept drift. Each histogram uses 40 bins.
• Figure 2: PHT for a dataset with two features. Each feature is represented over a period of 20 disjoint time windows. For each time window, we represent its histogram and the associated mean. The segments that join these successive means indicate the presence of drift. We observe that the drift is more pronounced in the first feature.
• Figure 3: Representation of the SINE1 dataset for 19 consecutive windows of 5,200 samples. The per class histograms and the means indicate the presence of sudden drift in both features.
• Figure 4: PHT representation for consecutive windows for the SINE1 dataset. The window size in (a) was chosen such that the concept drift changes are present inside the middle of a window and mimics the real scenario where drift position is not known apriori. In (b), the drift was identified by the SVM and we were able to isolate it in-between windows.
• Figure 5: PHT for the Forest Covertype dataset. Six of the ten continuous features were used. We show here the first 10 consecutive windows of 500 samples each. As the diagram shows, steady drift is present in the features.