Mining Spatio-Temporal Reachable Regions over Massive Trajectory Data Guojun Wu#, Yichen Ding#, Yanhua Li#, Jie Bao†, Yu Zheng†, Jun Luo*,$ #Worcester Polytechnic Institute (WPI) †Microsoft Research *Shenzhen Institutes of Advanced Technologies $Lenovo Research at Hong Kong 1

Imagine This… Location-based Advertising Aim to arrange special activities for business and need to decide spatial regions to distribute coupons The reachable region is time-varying 6 p.m. 1 p.m. In today’s talk, I am focusing on managing and analyzing big trajectory data from urban networks. In an urban network, vehicles, mobile phone users serve as urban roving sensors delivering a big trajectory data. These trajectories characterize versatile moving patterns, and reflect the urban issues, such as traffic jam, malicious events, etc. location-based advertising, where some business owner finds out the potential spatial regions to arrange special activities, such as distributing coupons and sales discount Orange Road means traffic jam, while green road means clear. 2

Real-world Problems ( a ) Location-based Recommendation ( b ) Location-based Advertising (a) A shopping mall distributes its special activities within reachable area. (b) Friends arrange their entertainment activities as much as possible. (c) Two delivery stations manage their affiliated delivery sites and dispatch ( c ) Business Coverage Analysis ( d ) Emergency Response Analysis

State-of-the-art works Related Work 10 miles 15 min 10 miles ST Reachability Our work Free space: circle 10 mile radius Road Network: road length 10 mile in all directions, with different time periods ST-Method: spatio-temporal reachable region, consider time aspect, with a given time period Free Space Road Network State-of-the-art works

Spatio-Temporal Reachability Query Definition: Find the reachable area in a spatial network from a location in a given time period. Example: Start from my home at 8:00 AM, where I can reach in 30 minutes One starting location Multiple starting locations

Naïve Method: Exhaustive Search Start from the querying location S and time T, then search the neighboring road segments throughout the whole road network.

Query Processing Framework

Stage 1: Preprocessing Road Re-segmentation Map Matching Partition the original road segments based on the given spatial granularity Map Matching Map trajectory point to real road network Chop based on intersection Road segment length differs a lot, 1mile to several feet. Map points to where it should be, based on distance and direction.

Stage 2: Indexing Structure ST-Index B-tree temporal R-tree spatial Nodetime list (different dates) item in time listdisk page

Stage 2: Indexing Structure Connection Index Based on speed information from trajectory data, we have two list for one segment ,at least where you can get starting from this segment and at most where you can get.

Stage 3: Query Processing Single Location Maximum bounding region search Trace back search With the ST-Index and Connection Index, now we are able to process queries from the user. In this processing, we employ 2 main techniques. One is the maximum/minimum bounding region search which use our index structures to generate a rough estimation of the upper bound of the reachable region based on the query parameters. The other is the Trace Back Search which uses connection index to refine the region from the rough bounding region. For the single location spatio-temporal reachability queries, we first estimate the maximum bounding region and then trace back search starts from the maximum bounding region to minimum one until satisfy the probability requirement. In this figure, the 2 solid circles indicate the maximum and minimum bounding region, respectively. And the dashed circle represents the prob-reachable region with respect to Probability requirement. As the result, all road segments within this dashed circle is to answer the single location query. maximum bounding region

Stage 3: Query Processing Multiple Locations Unified maximum bounding region search Trace back search r1 r2 r2 r4 r3 Compared with single location, multiple locations query has multiple starting locations. The difference in query processing is to estimate the unified maximum bounding region to skip the overlapping region search. As shown in left figure, road segment r1 and r2 are starting locations. Solid line indicates the outer boundary of the unified maximum bounding region of both r1 and r2 while dashed line indicates the boundary of overlapping region. And r3 is on the boundary of r2 while r4 on the r1. However, r3 is not on the outer-most boundary while r4 is. The right figure indicates the trace back search which is similar to single location starts from maximum bounding region to inner minimum one. r1 Unified bounding region Unified maximum bounding region

Evaluation Dataset: 194 GB real human mobility data in a large city in eastern China Baseline Algorithm Exhaustive search Evaluation metric Processing Time & Reachable Road Length Statistics Value City Size 400 square miles City Population Size three million people Duration 30 days in November, 2014 Number of taxis 21,385 unique taxis Number of trajectories 400 million (407, 040, 083)

Evaluation Effects on query probability Prob Exhaustive search and duration L = {10,15} min Processing Time vs. Reachable Road Length Decrease 50% Time

Visualization (a) Prob = 20% (b) Prob = 60% (c) Prob = 80% (d) Prob = 100%

Evaluation Effects on Start Time T Duration L = {5,10} min Processing Time vs. Reachable Road Length Same trend => consistent with each other

Visualization

Evaluation Comparison of s-query and m-query Duration L = {5,10,…,35}min Number of locations # = {1,2,…,9} Up to 70% Time Up to 90% Time

Visualization

Summary Spatio-Temporal Reachability Query Index Construction Spatio-Temporal Index Connection Index Query Processing Single-location ST Reachability Query via SQMB algorithm + TBS algorithm Multi-location ST Reachability Query via MQMB algorithm + TBS algorithm

Thank you! Contact: Yanhua Li (yli15@wpi.edu) Website: http://wpi.edu/~yli15/