A human activity recognition framework for grossly labeled smartphone sensing data through combining genetic algorithm with multiple instance multiple label learning

Human Activity Recognition through smartphones plays a crucial role in several medical state-of-affairs like patient monitoring, eldercare, and post-surgery recovery. Most of these works require precisely labeled accelerometer data for training supervised learning classifiers. Precise labeling of smartphone sensing data is difficult in real life due to the non-uniform gait of individual users with unpredictable activity transitions in between. Selecting an optimal number of features for the classification of such grossly labeled smartphone accelerometer instances is hardly investigated in the literature. Hence, a semi-supervised learning approach, that combines Multi-Instance Multi-label (MIML) learning with a Genetic Algorithm (GA) is proposed here. Rather than labeling a single instance, activity bags are designed for classification. GA selects the optimal sets of features for such grossly labeled bags of featured instances. MIML-kNN is chosen to be the classifier that is integrated with GA to predict the single, double or triple activity combinations that are often performed in daily life. Interestingly, the proposed framework can also predict component activities from an unknown activity combination even when it is not trained with such combinations. The framework is implemented for a real dataset collected from 8 users and it is found to be working adequately with less than half the set of features giving an average precision of 94% for even triple and double activity combinations.