Semi-supervised learning based activity recognition from sensor data

Abstract

The semi-supervised kernel logistic regression (SSKLR), developed for the classification of human behaviors from sensor data, takes the form of a linear combination of kernel functions associated with each of the labeled and unlabeled data from the training set. Its model parameters are determined, using an EM algorithm, by maximizing the expectation of the joint distribution over the posterior for selected unlabeled data that are in a neighborhood of one of labeled data. Tests for two types of human behaviors such as (1) "walk," and "skip," and (2) "drink a cup of tea," and "wash a cup" reveal that, using acceleration data as input, SSKLR classifies the behaviors better than semi-supervised Gaussian mixture and semi-supervised support vector machine models.