Collective Traffic Prediction with Partially Observed Traffic History using Location-Based Social Media

Transcription

1 Collective Traffic Prediction with Partially Observed Traffic History using Location-Based Social Media Xinyue Liu, Xiangnan Kong, Yanhua Li Worcester Polytechnic Institute February 22, / 34

2 About me 2 / 34

3 About me 3 / 34

4 About me I only know Python (2), and it is great. I think JavaScript, Ruby, Haskell... are cool, but I am too lazy to learn them. I hate C++. 4 / 34

5 My Research Interests Social Network Analysis [CIKM 16, SDM 17b] Recommender Systems [SDM 16] Brain Network [SDM 17a, IJCNN17] 5 / 34

6 Overview 1 Motivation 2 Related Works 3 CTP Method 4 Experiments 5 Summary 6 / 34

7 Why Traffic Prediction? Excessive traffic causes travel delays, resource wasting, and pollution. In 2011, traffic congestion costs urban Americans 5.5 billion hours of travel delay, 2.9 billion gallons of extra fuel, for a total congestion cost of $121 billion. 7 / 34

8 Why (Location-Based) Social Media? Location Associations Traffic jam on Storrow Drive, Boston, Massachusetts Temporal Data 8 AM 4 PM 11 PM Semantic Data Traffic Condition Sensor Traffic Networks Location-Based Social Media Location-Based Social Media (LBSM) is popular, can be used as mobile sensors. Semantic and spatial information from social media can be helpful. 8 / 34

9 Challenges Lack of historical traffic data in partial regions. In real-world road systems, only a small fraction of the road segments are deployed with sensors. It is difficult to predict traffic without traffic history. Sparsity of LBSM information at fine granularity. Table: Average # of tweets in each region under different spatiotemporal resolutions Temporal Resolution Spatial Resolution Ave. #Tweets 12 hours ,113 1 hour 1 1 3,926 1 hour 2 2 1,306 1 hour hour hour / 34

10 Conventional Methods Auto Regression [Smith and Demetsky, 1997, Journal of Transportation Engineering] Tweet Semantics [He et al.,2013, IJCAI] 10 / 34

11 Auto Regression [Smith and Demetsky, 1997] Prediction spatio-temporal dependencies Historical Traffic Data t time v (t) g = α + β 1 v (t 1) g + β 2 v (t 2) g Fail to work for locations without traffic history. 11 / 34

12 Tweet Semantics [He et al.,2013] Social Media Prediction e a c b d e a c b d e a c Historical Traffic Data t time Consider each location independently. Extract tweet semantics as bag-of-words feature for each location during a 12-hour time window. Build an auto regression-like model using both traffic history and tweet semantics. Fail to work for locations without traffic history. 12 / 34

13 Illustration of CTP [Our Method] road network a b c d e Local-based Social Media time a c b congestion Prediction spatio-temporal dependencies e d regions without any sensor Historical Traffic Data t time Incorporate LBSM information at finer spatiotemporal granularity. Consider different locations collectively. It works for locations without traffic history! 13 / 34

14 Social Media Semantic Vectors 14 / 34

15 Spatio-temporal Dependencies: I v i (t 1) t-1 v j (t 1) v q (t 1) v p (t 1) v i (t ) t v j (t ) v q (t ) v p (t ) Same as the traffic history in auto regression model. 15 / 34

16 Spatio-temporal Dependencies: II v i (t 1) t-1 v j (t 1) v q (t 1) v p (t 1) v i (t ) t v j (t ) v q (t ) v p (t ) Spatial dependency within a time window. 16 / 34

17 Spatio-temporal Dependencies: III v i (t 1) t-1 v j (t 1) v q (t 1) v p (t 1) v i (t ) t v j (t ) v q (t ) v p (t ) Spatial dependency across time windows. 17 / 34

18 CTP Method t-2 t-1 Training v " v & v ' Response LBSM Semantics assume time lag = 2 for the simplicity here. response variable (average speed, total traffic flow, etc). 18 / 34

19 CTP Method t-2 t-1 t-2 t-1 Training v " v & v ' Response LBSM Semantics Dependency I (Traffic History) 19 / 34

20 CTP Method t-2 t-1 t-2 t-1 v " ($%&) v " ($%() v ) ($%&) v ) ($%() v * ($%&) v * ($%() Training v " v ) v * Retrieve the historical data Response LBSM Semantics Dependency I (Traffic History) 20 / 34

21 CTP Method t-2 t-1 t-2 t-1 t Training v " v & v ' Response LBSM Semantics Dependency I (Traffic History) Dependency II (Neighbors Traffic) 21 / 34

22 CTP Method t-2 t-1 t-2 t-1 t t-2 t-1 Training v " v & v ' Response LBSM Semantics Dependency I (Traffic History) Dependency II (Neighbors Traffic) Dependency III (Neighbors Traffic History) 22 / 34

23 CTP Method t-2 t-1 t-2 t-1 t t-2 t-1 Training v 6 v " v * Compute using an aggregation function (e.g. average) Response LBSM Semantics Dependency I Dependency II Response = Speed, aggregation function = AVG. v " = 50, v* = 45, v, and v- are unobserved. The Dependency-II Feature for node A at time t is: (/ (1) (1) 0 + / 2 ) = Dependency III 23 / 34

24 CTP Method t-2 t-1 t-2 t-1 t t-2 t-1 Training (only observed) Response LBSM Semantics Dependency I Dependency II t-1 t t-1 t t+1 t-1 t Bootstrap (unobserved regions) Dependency III 24 / 34

25 CTP Method t-2 t-1 t-2 t-1 t t-2 t-1 Training (only observed) Response LBSM Semantics Dependency I Dependency II t-1 t t-1 t t+1 t-1 t Bootstrap (unobserved regions) Dependency III 25 / 34

26 CTP Method t-2 t-1 t-2 t-1 t t-2 t-1 Training (only observed) t-1 t t-1 t t+1 t-1 t Iterative Inference Response Keep updating LBSM Semantics Dependency I 0 0 Keep updating (unobserved regions) Dependency II Dependency III 26 / 34

27 Dataset Traffic Data Collect from the California Performance Measurement System(PeMS) between October 19 and November 28, ,102,272 entries of traffic records. LBSM Data Collect tweets from the same area during the same time range using the Twitter streaming API. This collection results in a total number of 2,648,446 tweets. 27 / 34

28 Compared Methods TDO[Smith and Demetsky, 1997]: Auto regression model using traffic history. TDO-floor[ ]: Similar to TDO, except it uses full traffic history. TwSeO: A degenerated version of [He et al. 2013], using tweets semantics. 28 / 34

29 Experimental Setting Partition the data into two parts, with the beginning (1 1 u ) as the training set and the remaining 1 u as the test set (u = 3,..., 7). k-fold cross-validation is used to randomly sample 1/k regions as unobserved (k = 2, 3, 4, 5). Root Mean Square Error (RMSE) is used to evaluate the performance. 29 / 34

30 Results our method lower Is better TDO-floor performs the best by using full traffic history. The proposed CTP outperforms TDO and TwSeO. The result shows the effectiveness of incorporating tweets semantics into the collective inference model. 30 / 34

31 The effect of r lower Is better our method Sparser Information in LBSM Figure: Test Ratio = 1/7 (u = 7) 31 / 34

32 The effect of k our method lower Is better Less Unobserved Regions Figure: u = 6, r = 5 32 / 34

33 Summary Problem Studied Traffic prediction with partially observed traffic history. Proposed Model Using LBSM data to alleviate the issue of absent traffic history. A collective inference model that exploits the complex spatio-temporal dependencies between road segments as well as incorporates LBSM semantics in the prediction. 33 / 34

34 Q&A Xinyue Liu Xiangnan Kong Yanhua Li 34 / 34