1. Traffic Prediction with Partially Observed History Collective Traffic Prediction with Partially Observed Traffic History using Location-Based Social Media Xinyue Liu, Xiangnan Kong and Yanhua Li Worcester Polytechnic Institute Changing the Template 2. Location-Based Social Media (LBSM) 3. Collective Inference Framework Changing the Template 4. Experiment Results fig_result2_rmse_split4.pdf 5. Acknowledgement This work is supported in part by the National Science Foundation through grant CNS-1626236. t time Prediction spatio-temporal dependencies Historical Traffic Data road network a b c e d regions without any sensor 1. Problem Studied: Traffic Prediction with Partially Observed Traffic History. 2. Traffic Prediction: Infer the traffic conditions in the future time span for a geographical area. 3. Partially Observed History: In the target area, some locations are not deployed with any sensors, their historical traffic conditions are not available. 4. The Goal: Make good predictions for every location in the target area based on the partially observed traffic history. t-1 t v i (t −1) v j (t −1) v q (t −1) v p (t −1) v i ( t ) v j ( t ) v p ( t ) v q ( t ) t-1 t v i (t −1) v j (t −1) v q (t −1) v p (t −1) v i ( t ) v j ( t ) v p ( t ) v q ( t ) t-1 t v i (t −1) v j (t −1) v q (t −1) v p (t −1) v i ( t ) v j ( t ) v p ( t ) v q ( t ) Why LBSM ? Challenge 0: How to incorporate LBSM semantics into traffic prediction? Challenge 1: Lack of Traffic History Information for Some Locations. Challenge 2: Sparsity of LBSM Information at Fine Granularities (Table 1). 1. Cover much wider range of geographic areas. 2. Provide abundant information about the road users in real-time. 3. Dictation Systems (e.g. Siri) in smart phones or smart cars allow road user to post contents in LBSM easily . 4. By mining the semantic and spatial information from LBSM, we can effectively infer the future traffic conditions for many areas, including the road segments without sensors. Target Area: Greater Los Angeles LBSM Used: Twitter Bag-of-words Representation Stemmed Words Traffic History Inter-Region History Neighboring Traffic t time Prediction Social Media Historical Traffic Data a c e a b c e d a b c e d LBSM Semantics Traffic History Neighboring Traffic Inter-Region History t-2 t-1 t-1 t-2 t t-2 t-1 0 0 0 0 Training … Bootstrap Iterative Inference 0 0 (unobserved region) (unobserved region) Response Keep updating Keep updating t-1 t t t-1 t+1 t-1 t (only observed) (only observed) (observed region) (observed region) … … 1/k regions are unobserved 1/3 Test Data 20 X 20 Grids 1/k regions are unobserved 1/k regions are unobserved 1/k regions are unobserved 1/6 Test Data 5 X 5 Grids 1/6 Test Data 20 X 20 Grids 1/4 Test Data 1/5 Test Data 1/6 Test Data 1/7 Test Data 1/3 Test Data 5 X 5 Grids J. He et al. Auto- Regression t time Prediction congestion spatio-temporal dependencies time a b c d e Local-based Social Media Historical Traffic Data Our Model