University of Oulu

Siirtola, P.; Röning, J. Incremental Learning to Personalize Human Activity Recognition Models: The Importance of Human AI Collaboration. Sensors 2019, 19, 5151

Incremental learning to personalize human activity recognition models : the importance of human AI collaboration

Saved in:
Author: Siirtola, Pekka1; Röning, Juha1
Organizations: 1Biomimetics and Intelligent Systems Group, University of Oulu, P.O. BOX 4500, FI-90014 Oulu, Finland
Format: article
Version: published version
Access: open
Online Access: PDF Full Text (PDF, 0.6 MB)
Persistent link:
Language: English
Published: Multidisciplinary Digital Publishing Institute, 2019
Publish Date: 2020-01-07


This study presents incremental learning based methods to personalize human activity recognition models. Initially, a user-independent model is used in the recognition process. When a new user starts to use the human activity recognition application, personal streaming data can be gathered. Of course, this data does not have labels. However, there are three different ways to obtain this data: non-supervised, semi-supervised, and supervised. The non-supervised approach relies purely on predicted labels, the supervised approach uses only human intelligence to label the data, and the proposed method for semi-supervised learning is a combination of these two: It uses artificial intelligence (AI) in most cases to label the data but in uncertain cases it relies on human intelligence. After labels are obtained, the personalization process continues by using the streaming data and these labels to update the incremental learning based model, which in this case is Learn++. Learn++ is an ensemble method that can use any classifier as a base classifier, and this study compares three base classifiers: linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), and classification and regression tree (CART). Moreover, three datasets are used in the experiment to show how well the presented method generalizes on different datasets. The results show that personalized models are much more accurate than user-independent models. On average, the recognition rates are: 87.0% using the user-independent model, 89.1% using the non-supervised personalization approach, 94.0% using the semi-supervised personalization approach, and 96.5% using the supervised personalization approach. This means that by relying on predicted labels with high confidence, and asking the user to label only uncertain observations (6.6% of the observations when using LDA, 7.7% when using QDA, and 18.3% using CART), almost as low error rates can be achieved as by using the supervised approach, in which labeling is fully based on human intelligence.

see all

Series: Sensors
ISSN: 1424-8220
ISSN-E: 1424-8220
ISSN-L: 1424-8220
Volume: 19
Issue: 23
Article number: 5151
DOI: 10.3390/s19235151
Type of Publication: A1 Journal article – refereed
Field of Science: 213 Electronic, automation and communications engineering, electronics
Funding: This research is supported by the Business Finland funding for Reboot IoT Factory-project (
Copyright information: © 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (