Mobility Event Detection And Prediction Based On Mobile Data

With the continuous development of mobile positioning devices,a large number of spatio-temporal data,including Call Detail Records(CDRs),GPS and WiFi,are generated and provide a powerful support for crowds flow anal-ysis,pushing forward the rapid development of crowds flow prediction[1-4]and local event detection[5-8].And These crowds flow analysis at urban level can assist in monitoring the risks of citywide security to some extent.How-ever,there are still some drawbacks in current studies.Current studies on crowds flow prediction generally weaken the complex relationship between spatio-temporal crowds flow data and external factors.In addition,the mo-bility pattern information,such as popular travel and commuting patterns,were generally neglected.And as for the current studies on local event detection,the crowds flow and social media data were not fully fused and there were also some limitations in the mining of anomaly feature patterns.Therefore,we tend to improve the crowds flow prediction and local event detection framework based on mobile big data.The main work and innovation of this paper are listed as follows:(1)Propose a multi-view,including spatio-temporal and mobility pattern view,crowds flow prediction framework MV-RANet.With the fully fusion of mobility pattern information and the fine-grained modeling of the complex de-pendencies between spatio-temporal properties and external factors based on multi-channel and residual attention mechanism,the crowds flow prediction framework was further improved.Experiments on two real-world datasets,CDR and TaxiBJ[9],demonstrated the efficiency of our proposed MV-RANet.(2)Propose a local event detection framework,ED-BiGAN,based on cross-domain data.With the fully fused features extracted from mobile trajectory and social media data based on multi-channel mechanism,the local event detection was implemented by Bidirectional Generative Adversarial Networks(BiGAN).Experimental results on two real world event detection datasets[10,11]demon-strated that our ED-BiGAN significantly outperforms all the baselines.(3)Propose a fusion method that incorporates the local event informa-tion into the crowds flow prediction framework.With the residual attention fusion mechanism,the performance of the crowds flow prediction was further improved,especially the robustness of the framework when some local event took place.