UDetect: Unsupervised Concept Change Detection for Mobile Activity Recognition

Abstract

One of the major challenges in activity recognition task is the need to adapt a classification model during its operation. This is important because the underlying data distribution between those used for training and the new evolving stream of data may change during online recognition. The changes between the two sessions may occur because of differences in sensor placement, orientation and user characteristics such as age and gender. However, many of the existing approaches for model adaptation in activity recognition are blind methods because they continuously adapt the classification model without explicit detection of changes in the concepts being predicted. Therefore, we propose a concept change detection method for activity recognition under the assumption that a concept change in the model of an activity is followed by changes in the distribution of the input data attributes as well which is the realistic case for activity recognition. Our change detection method computes change detection statistic on stream of multi-dimensional unlabelled data that are classified into different concept windows. The values of the change indicators are then processed for detecting peak points that indicate concept change in the stream of activity data. Evaluation of the approach using real activity recognition dataset shows consistent detections that correlate with the error rate of the model.