Study of similarity metrics for matching network-based personalised human activity recognition.

Personalised Human Activity Recognition (HAR) models trained using data from the target user (subject-dependent) have been shown to be superior to non personalised models that are trained on data from a general population (subject-independent). However, from a practical perspective, collecting sufficient training data from end users to create subject-dependent models is not feasible. We have previously introduced an approach based on Matching networks which has proved effective for training personalised HAR models while requiring very little data from the end user. Matching networks perform nearest-neighbour classification by reusing the class label of the most similar instances in a provided support set, which makes them very relevant to case-based reasoning. A key advantage of matching networks is that they use metric learning to produce feature embeddings or representations that maximise classification accuracy, given a chosen similarity metric. However, to the best of our knowledge, no study has been provided into the performance of different similarity metrics for matching networks. In this paper, we present a study of five different similarity metrics: Euclidean, Manhattan, Dot Product, Cosine and Jaccard, for personalised HAR. Our evaluation shows that substantial differences in performance are achieved using different metrics, with Cosine and Jaccard producing the best performance.