One-Class Classifier-based Incremental Learning Method to Personalize Multi-Class Human Activity Recognition Models from Streaming Data

Abstract

Human-activity-recognition (HAR) systems trained on population-level data often provide lower accuracy when used with new users. This article introduces an incremental learning-based method to personalize a multi-class HAR model using only the streaming accelerometer data naturally produced during everyday use. In this article, incremental learning is based on ensemble method called Learn++. The core idea of the proposed method is to replace the traditional multi-class base classifier in Learn++ with one-class classifiers, own one-class classifier for each activity. Because every class is modelled in isolation, the ensemble can be updated whenever labelled data for just one activity is available. Therefore, in order to update model, there is no need to collect a multi-class chunk of data and retrain the model like usually. Approach is evaluated with accelerometer data from 24 study subjects containing data from five activities. With Isolation Forest as the base learner, mean accuracy rises from 76.4% to 84.5%, and for 20 of 24 users the recognition rate improves. In some cases this improvement is as high as 20 percentage units. Activity-wise analysis shows the greatest improvements for non-static activities such as Walking and Stairs. The method supports frequent updates, accommodates new activities by simply adding further one-class models, and runs without storing large datasets - making it suitable for resource-constrained wearables.