Learning to Compare with Few Data for Personalised Human Activity Recognition

Recent advances in meta-learning provides interesting opportunities for CBR research, in similarity learning, case comparison and personalised recommendations. Rather than learning a single model for a specific task, meta-learners adopt a generalist view of learning-to-learn, such that models are rapidly transferable to related (but different) new tasks. Unlike task-specific model training, a meta-learner’s training instance - referred to as a meta-instance - is a composite of two sets: a support set and a query set of instances. In our work, we introduce learning-to-learn personalised models from few data. We motivate our contribution through an application where personalisation plays an important role, mainly that of human activity recognition for self-management of chronic diseases. We extend the meta-instance creation process where random sampling of support and query sets is carried out on a reduced sample conditioned by a domain-specific attribute; namely the person or user, in order to create meta-instances for personalised HAR. Our meta-learning for personalisation is compared with several state-of-the-art meta-learning strategies: 1) matching network (MN), which learns an embedding for a metric function; 2) relation network (RN) that learns to predict similarity between paired instances; and 3) MAML, a model-agnostic machine-learning algorithm that optimizes the model parameters for rapid adaptation. Results confirm that personalised meta-learning significantly improves performance over non personalised meta-learners.