SIGSPATIAL Cup
Papers of Winning Teams (presentation in reverse order)
- Complementary Fusion of Deep Network and Tree Model for ETA Prediction
YuRui Huang (Nanjing University of Science and Technology), Jie Zhang (Huatai Securities), Hengda Bao (Baidu), Jian Yang (Nanjing University of Science and Technology), Yang Yang (Nanjing University of Science and Technology) - Travel Time Estimation Based on Neural Network with Auxiliary Loss
Yunchong Gan (Peking University), Haoyu Zhang (Peking University), Mingjie Wang (Beijing Normal University) - Geo-Attention Network for Traffic Condition Prediction and Travel Time Estimation
Jie-Zhang Li, Wangyi Zhou, Zebin Chen, Yue-Jiao Gong (South China University of Technology) - Multi-View Spatial-Temporal Model for Travel Time Estimation
Zichuan Liu (Wuhan University of Technology), Zhaoyang Wu (East China Normal University), Meng Wang (Sun Yat-sen University) - Estimated Time of Arrival Prediction via Modeling the Spatial-Temporal Interactions between Links and Crosses
Xiaowei Mao, Tianyue Cai, Wenchuang Peng, Huaiyu Wan (Beijing Jiaotong University, Beijing Key Lab of Traffic Data Analysis and Mining) - Hierarchical Positional Approach for ETA Prediction
Tomoki Saito, Shinichi Tanimoto, Fumihiko Takahashi (Mobility Technologies Co., Ltd.) - Integration Model for Estimated Time of Arrival
Xuewei Guo, Shenglong Zhang (AI Labs, yz-intelligence Inc.)
This year's contest focuses on estimated time of arrival (ETA), aka travel time prediction of a trip on a road network. ETA plays an important role in various applications, for instance, in ride-hailing scenes like order dispatch, price estimation, arrival time prediction, route decision, etc. Accurate estimation could improve user experience and system efficiency for the service platform, thus helping both users and the platform make better decisions. However, travel time may be affected by route distance, road capacity, road level, real-time traffic, traffic light and other traffic elements. Among these elements, route distance, link info, and traffic light info are static, but real-time traffic conditions are always dynamic. Moreover, travel time has strong periodicity. Even within the same route, travel time may vary a lot between weekdays and weekends, flat periods and peak periods. In summary, ETA has been a challenging problem for a long time with high technical difficulty.
The difficulties of ETA problems lie in:
- Differences of driver behaviors;
- Unstable average through time in the areas of traffic lights;
- Low accuracy in sum of average time due to cumulative error;
- Variations of traffic conditions during the trip, especially in peak periods, since only that of the departure time can be obtained;
- Pronounced differences in travel time between weekdays and weekends, as well as between peak and off-peak periods.
The key topics that are being addressed in this year’s contest are the following:
- Spatio-temporal Big Data
- Spatial Networks
- Travel Time Prediction
One peculiarity of this year’s contest is the interactive online submission mode that allows each contestant team to submit their prediction results up to 5 times every day and see their most up-to-date rankings. Results need to be submitted on Biendata https://biendata.xyz/competition/didi-eta/. Five top teams that rank best on the Leaderboard will be provided with cash and/or other prizes. In addition to these prizes, these top five teams will be invited to submit a four-page paper for a contest paper session to be held at the 2021 ACM SIGSPATIAL GIS conference. These papers will be subject to review and acceptance by the contest organizers, but it is expected that each of the top five teams will have their paper in the conference proceedings and a ten-minute presentation in the contest session. At least one team member of each winning team must register for the 2021 ACM SIGSPATIAL GIS conference.
Problem Definition
The motivation of this problem is prediction of the travel time, given the departure time, route info, real-time traffic and weather features.
Input:
- A road network (map)
- A training dataset (trip info from August 1st to August 31th, Shenzhen in China, organized by day)
- A test dataset (trip info of Sep 1st, Shenzhen in China)
Output:
Estimated time of arrival of the testing data.
Objective:
Minimizing the mean absolute percentage error between the estimated time of arrival and the actual travel time.
Data Description:
Trip trajectories and real-time traffic conditions
| Part | Field | Type | Description |
| head part | order id | string | unique order id |
| ata | float | travel time | |
| distance | float | route distance | |
| simple eta | float | accumulation of average link time of departure time | |
| driver id | int | unique driver id | |
| slice id | int | time slice of departure time | |
| link part | link id | int | road segment id |
| link time | float | average through time of departure time | |
| link ratio | float | road segment through ratio | |
| link current status | int | link traffic condition of departure time | |
| link arrival status | int | link traffic condition of arrival time | |
| cross part | cross id | int | traffic light cross id |
| cross time | float | mining cross time of traffic light |
Fields descriptions:
- Delimiter of different part is ";;", and the fields are split by space delimiter.
- Trip trajectories are composed of link part and cross part. These parts are organized in sequence, every element of the sequence is composed by the data fields in the table.
- Simple eta are the accumulation of the link time and cross time.
- Slice id is the time bucket of the departure time, the bucket size is every 5 minutes.
- Link time is the average through time of 10 minutes before.
- Link ratio is the coverage ratio of the real trajectory and the road segment. In the begging and the end of the trip, link ratio will be less than 1.0.
- Link current status is the real time traffic condition of the road segment in the departure time. Traffic condition is expressed by 1 (unblocked), 2 (slow), 3 (jam) and 0 (unknown).
- Link arrival status is the real time traffic condition of the road segment in arrival time. PS. this feature is missing in test data.
- cross id is the unique id of a traffic light, which is composed of in link and out link of the cross, cross time is the mining waiting time of this traffic light.
- All the trip data is organized by day. Each day is a separate file.
Sketch Map：
Road network topology
| key | value |
| linkid | next link id1, next link id2, ... |
Contest Registration & Data Download
Follow the instructions on https://www.biendata.xyz/competition/didi-eta/ for contest registration and data download.
Important Dates
Deadline of submission: August 9th, 2021. (23:59, UTC) Notification of the result: August 30th, 2021. (23:59 UTC) Submission deadline of invited papers: September 15th, 2021. (23:59 UTC)
Submission and Evaluation
The Cup participants shall submit their results on biendata. https://biendata.xyz/competition/didi-eta/
Cash Rewards
Total bonus: $25,000 1st Place: $10,000 2th to 3th Place: $5,000 for each 4th to 5th Place: $2,500 for each ⋆The prizes are before tax.
Contest Chairs
Hua Chai, Didi Chuxing, China Dimitris Sacharidis, Vienna University of Technology, Austria Fang Jin, George Washington University, USA
Sponsorship
This edition of ACM SIGSPATIAL GIS CUP is sponsord by:
Previous Editions
https://sigspatial2020.sigspatial.org/giscup2020/home https://sigspatial2019.sigspatial.org/giscup2019/home https://sigspatial2018.sigspatial.org/giscup2018/home https://sigspatial2017.sigspatial.org/giscup2017/home