Abstract
"The developments in data collection technologies and their growing availability lead to an extensive amount of time-series data. The collected data contain information from people's day-to-day lives, health status or activities. However, this huge amount of data need to be analysed to extract useful information. These insights about human lives and health can be helpful in different fields and areas such as marketing, transportation, health tracking and improving quality of life, education and safety.
There are extensive research and studies to improve the time-series analysis. One area which has gained more attention is machine learning and deep learning. Machine learning can be useful in gathering insight from massive datasets and developing models for predictive reasons. However, there are challenges in applying machine learning methods and techniques to time-series data.
This research proposes a number of methods to tackle some of the challenges and issues in this area. This work proposes a novel representation model to deal with high-dimensionality of time-series data. The model preserves the principal information of data and we demonstrate how this data representation can be used in different time-series data analysis such as clustering, classification and change point detection. We also compared the proposed model with other representation models in clustering which showed 20% higher clustering quality in Silhouette coefficient measure.
Healthcare and especially remote healthcare is an area that can benefit from time-series data analysis. An attention-based model has also been implemented in a dementia care project. Attention-based model is a deep learning model that can find important features in data and give higher weight to them in the prediction. The proposed model is able to detect the risk of adverse health related incidents at the homes of people with dementia with a recall of 91% and precision of 83%.
Finally, to further improve robustness of the analysis methods, a semi-supervised model has been developed. The model can use a limited amount of labelled data and improve the performance compared to other state-of-the-art models. The model shows 27% higher recall in average compared with other models in detecting agitation in people with dementia."