Vers des services Internet basés sur les profils de mobilité des utilisateurs

Abstract

Nowadays, mobility prediction models play an important role in many locationbased services, such as food delivery, transportation planning, and advertisement posting. Most previous studies on predicting mobility have worked on computer generated data and focused on mathematical modeling principally due to the lack of a real mobility data. Such studies have limited ability to capture human mobility accurately. However, with the democratization of mobility data and the availability of large data sets, numerous research activities turned toward predicting mobility based on examining real mobility data traces with the aim of building realistic models that can capture and understand human’s mobility behaviors as well as making accurate mobility prediction. In this thesis, we present the methods we proposed to predict spatial and temporal behaviors of mobile users. Our first work focuses on predicting the next location of mobile users by analyzing large data sets of the history of their movements. We make use of past location sequences, also called location history, to train a classification model that will be used to predict future locations. Contrary to traditional mobility prediction techniques based on Markovian models, we investigate the use of modern deep learning techniques such as the use of Convolutional Neural Networks (CNNs). Inspired by the word2vec embedding technique used for the next word prediction, we present a new method called loc2vec in which each location is encoded as a vector whereby the more often two locations cooccur in the location sequences, the closer their vectors will be. Using the vector representation, we divide long mobility sequences into several sub-sequences and use them to form Mobility Subsequence Matrices on which we run CNN classification which will be used later for the prediction. We run extensive testing and experimentation on a subset of a large real mobility trace database made publicly available through the CRAWDAD project. Our results show that loc2vec embedding and CNN-based prediction provide significant improvement in the next location prediction accuracy compared to state-of-the-art methods. We also show that transfer learning on existing pre-trained CNN models provides further improvement over CNN models build from scratch on mobility data. We also show that our loc2vec-CNN model enhanced with transfer learning achieves better results than other variants including our iother proposal onehot-CNN and existing Markovian models. In the second work, we focus on predicting the temporal behavior, particularly the residence time, of mobile users at their relevant locations. In this work, we explored the joint use of location history, arrival time, and the previous residence time to accurately predict the residence time at the current location. We developed a model that integrates all these parameters and uses our modified Moving-Average and CDF time-aided algorithms that include the arrival time in the model. We run performance evaluation experiments on a subset of the same mobility trace collected by Dartmouth College. Our results show that adding high-granularity temporal information to the mobility model allows to significantly improve the residence time prediction compared to state-of-theart methods. The prediction accuracy improvement for the dataset we work on has been consistent and of about 20% on the average. We also presented two linear mobility models for residence time prediction, namely Linear Regression (LR), and Auto-Regression (AR). We run performance evaluation experiments on two different WiFi mobility traces datasets made available through the CRAWDAD project. Our results show that using linear regression-based learning algorithms significantly improve the residence time prediction accuracy compared to stateof-the-art methods, and achieve prediction errors in the order of seconds and minutes for a large number of users