Inferring travel activity pattern from smartphone sensing data using deep learning

Abstract

Understanding the travel routine of the individuals is important in many domains. In transport research understanding daily travel routine is crucial for modeling the travel behavior of the individuals. Such models help predict the travel demand and develop strategies for managing that demand. Understanding travel patterns of the individuals is also important to develop effective incentive mechanisms. Location-based services like personal digital assistants and journey planners use historical travel routine to build preferences of the user and make useful recommendations. In health sciences logging the routine travel behavior is important to monitor health of the patients and make recommendations wherever necessary. Several fitness tracking applications available on smartphones utilize the travel activity diary to evaluate the fitness of the individuals and make recommendations. The proliferation of sensing-enabled smartphone devices engendered the development of tools for logging travel routine of individuals. The research in this thesis uses the sensor data collected from smartphone devices to develop a travel activity inference algorithm. Presently, the research into travel activity inference has been focused on developing supervised learning algorithms. These algorithms require a large amount of labeled data for training algorithms that generalize well. Generalization in personalized travel activity inference is a challenging problem due to the concept drift. The problem of concept drift is magnified as the more personalized information is introduced in the input variables. Once the users start using the applications they are constantly generating new data. Expecting the users to label all the data generated by them is impractical. Instead, it would be useful to identify only those examples which would help most improve the algorithm and have the user label such instance. This reduces the burden on the user and does not discourage them from participating in the data collection process. In other words, we need a model that is identifies concept drift in data and adapts accordingly. There has been advances in the deep learning research in last few years. The deep learning algorithms provide a framework for learning feature representation from raw data. The convolutional neural networks have been particularly effective in learning feature representations on many datasets. These models have achieved significant improvement on many complex problems over other machine learning approaches. For the sequential classification problems like the travel activity inference, the recurrent neural network like long short term memory networks are particularly suitable. This thesis proposes to use the deep learning algorithms for travel activity inference. To develop an end-to-end deep learning algorithm that learns feature representations from raw sensor data and incorporates different sensors with differing frequencies. The research proposes using a combination of convolutional neural network for feature representation learning in both time and frequency domain and long short term memory network for sequential classification. In practical situations, the users of the smartphones cannot be asked to carry their smartphones in a fixed position every time. The proposed algorithm for travel activity inference need to be robust to changes in orientation of the smartphones. We compared the performance of the proposed deep learning algorithm against a baseline model based on the current supervised machine learning approaches. The deep learning algorithm achieved an overall average accuracy of 95.98% compared to the baseline method which achieved an overall average accuracy of 89%. We also show that the proposed deep learning algorithm is robust to changes in the orientation of the smartphone.