Wearable biosensor network enabled multimodal daily-life emotion recognition employing reputation-driven imbalanced fuzzy classification

Daily-life emotion recognition is a new procedure developed from basic emotion recognition. It records and analyzes emotion-related bio-signals to evaluate emotional states of subjects when they are participating in daily tasks instead of receiving specific stimulations. This paper develops a wearable biosensor network to take a step further towards daily-life emotion recognition. Multimodal bio-signals (electroencephalography, pulse, skin temperature and blood pressure) are recorded by the sensor nodes and transmitted to the remote web data center through a body station to realize the web-enabled recognition scheme. In total, a 103-day emotion diary is kept from Jun 4th 2015 to Feb 28th 2016, discontinuously. The remarkably different appearing possibilities of 4 emotional states (horror, happiness, boredom and relaxation) and the noisy sensing environment create an imbalanced and noisy dataset. Thus, a reputation-driven imbalanced fuzzy support vector machine (RI-FSVM) classification method is proposed to reduce the adverse effects caused by both within-class noisy samples and between-class imbalance. The fuzzy membership function is determined by the reputation values (indicating the reliability of samples) and the class-imbalanced ratios. The experiment convinces that the wearable biosensor network works well and successfully extracts efficient features from multimodal bio-signals. These features are convinced to have better performance than the related work in both centrality and distinguishability. The proposed method improves the sensitivity, specificity and Gm of emotion classification compared with the typical classification methods. Eventually, our research achieves a competitive accuracy with a low-cost consumer-grade sensing system. The main contributions of this paper are the quantitative analysis on emotion diary and the imbalanced classification algorithm for daily-life emotion recognition.