Dynamic Ensemble Pruning Selection using Meta-Learning for Multi-Sensor Based Activity Recognition

• Published in: 2019 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI) pp. 1063 - 1068
• Main Authors: Cao, Jingjing, Yuan, Wenjing, Li, Wenfeng, E., Xiaozheng
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.08.2019
• Subjects: Activity recognition; Computational complexity; Computational modeling; Dynamic ensemble selection; Ensemble pruning; Feature extraction; Heuristic algorithms; Hidden Markov models; Human activity recognition; Meta-learning; Multi-sensor data fusion; Training
• DOI: 10.1109/SmartWorld-UIC-ATC-SCALCOM-IOP-SCI.2019.00204

With the ever-increasing sensor types and complexity in the field of activity recognition, proper multi-sensor configuration system is essential to balance the recognition performance improvement and increased computational complexity caused by the use of multiple homogeneous or heterogeneous sensors. The multi-sensor deployment problem is normally transformed to multiple ensemble classifier pruning problem, while the competence of various ensemble pruning approaches for a particular subject are generally different. In this paper, a dynamic ensemble pruning selection model using meta-learning (META-DEPS) is proposed to recommend the most competent ensemble pruning algorithm for a given test sample. We utilize statistics features of processed sensor dataset to form a representative meta-feature set. Then meta-features extracted from the training dataset are used to train the instance-based meta-learner with newly designed ranking rules. Finally, an ensemble pruning algorithm is selected based on the trained meta-learner and realize the prediction of test instance in the generalization phase. The system conducts empirical studies on real-world activity recognition dataset and the results show that the proposed meta-learning based framework improves the accuracy of activity recognition when compared against using conventional ensemble pruning.