Getting a good feature representation of data is paramount for Human Activity Recognition (HAR) using wearable sensors. An increasing number of feature learning approaches—in particular deep-learning based—have been proposed to extract an effective feature representation by analyzing large amounts of data. However, getting an objective interpretation of their performances faces two problems: the lack of a baseline evaluation setup, which makes a strict comparison between them impossible, and the insufficiency of implementation details, which can hinder their use. In this paper, we attempt to address both issues: we firstly propose an evaluation framework allowing a rigorous comparison of features extracted by different methods, and use it to carry out extensive experiments with state-of-the-art feature learning approaches. We then provide all the codes and implementation details to make both the reproduction of the results reported in this paper and the re-use of our framework easier for other researchers. Our studies carried out on the OPPORTUNITY and UniMiB-SHAR datasets highlight the effectiveness of hybrid deep-learning architectures involving convolutional and Long-Short-Term-Memory (LSTM) to obtain features characterising both short- and long-term time dependencies in the data.

The scarcity of labelled time-series data can hinder a proper training of deep learning models. This is especially relevant for the growing field of ubiquitous computing, where data coming from wearable devices have to be analysed using pattern recognition techniques to provide meaningful applications. To address this problem, we propose a transfer learning method based on attributing sensor modality labels to a large amount of time-series data collected from various application fields. Using these data, our method firstly trains a Deep Neural Network (DNN) that can learn general characteristics of time-series data, then transfers it to another DNN designed to solve a specific target problem. In addition, we propose a general architecture that can adapt the transferred DNN regardless of the sensors used in the target field making our approach in particular suitable for multichannel data. We test our method for two ubiquitous computing problems—Human Activity Recognition (HAR) and Emotion Recognition (ER)—and compare it a baseline training the DNN without using transfer learning. For HAR, we also introduce a new dataset, Cognitive Village-MSBand (CogAge), which contains data for 61 atomic activities acquired from three wearable devices (smartphone, smartwatch, and smartglasses). Our results show that our transfer learning approach outperforms the baseline for both HAR and ER.

Artificial intelligence and especially deep learning methods have achieved outstanding results for various applications in the past few years. Pain recognition is one of them, as various models have been proposed to replace the previous gold standard with an automated and objective assessment. While the accuracy of such models could be increased incrementally, the understandability and transparency of these systems have not been the main focus of the research community thus far. Thus, in this work, several outcomes and insights of explainable artificial intelligence applied to the electrodermal activity sensor data of the PainMonit and BioVid Heat Pain Database are presented. For this purpose, the importance of hand-crafted features is evaluated using recursive feature elimination based on impurity scores in Random Forest (RF) models. Additionally, Gradient-weighted class activation mapping is applied to highlight the most impactful features learned by deep learning models. Our studies highlight the following insights: (1) Very simple hand-crafted features can yield comparative performances to deep learning models for pain recognition, especially when properly selected with recursive feature elimination. Thus, the use of complex neural networks should be questioned in pain recognition, especially considering their computational costs; and (2) both traditional feature engineering and deep feature learning approaches rely on simple characteristics of the input time-series data to make their decision in the context of automated pain recognition.