Recent studies have shown the potential of exploiting GPS data for passively inferring people's mental health conditions. However, feature extraction for characterizing human mobility remains a heuristic process that relies on the domain knowledge of the condition under consideration. Moreover, we do not have guarantees that these "hand-crafted" metrics are able to effectively capture mobility behavior of users. Indeed, informative emerging patterns in the data might not be characterized by them. This is also a complex and often time-consuming task, since it usually consists of a lengthy trial-and-error process. In this paper, we investigate the potential of using autoencoders for automatically extracting features from the raw input data. Through a series of experiments we show the effectiveness of autoencoder-based features for predicting depressive states of individuals compared to "hand-crafted" ones. Our results show that automatically extracted features lead to an improvement of the performance of the prediction models, while, at the same time, reducing the complexity of the feature design task. Moreover, through an extensive experimental performance analysis, we demonstrate the optimal configuration of the key parameters at the basis of the proposed approach.
Using Unsupervised Deep Autoencoders to Automatically Extract Mobility Features for Predicting Depressive States
Musolesi, M
2018
Abstract
Recent studies have shown the potential of exploiting GPS data for passively inferring people's mental health conditions. However, feature extraction for characterizing human mobility remains a heuristic process that relies on the domain knowledge of the condition under consideration. Moreover, we do not have guarantees that these "hand-crafted" metrics are able to effectively capture mobility behavior of users. Indeed, informative emerging patterns in the data might not be characterized by them. This is also a complex and often time-consuming task, since it usually consists of a lengthy trial-and-error process. In this paper, we investigate the potential of using autoencoders for automatically extracting features from the raw input data. Through a series of experiments we show the effectiveness of autoencoder-based features for predicting depressive states of individuals compared to "hand-crafted" ones. Our results show that automatically extracted features lead to an improvement of the performance of the prediction models, while, at the same time, reducing the complexity of the feature design task. Moreover, through an extensive experimental performance analysis, we demonstrate the optimal configuration of the key parameters at the basis of the proposed approach.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
