SurvCancel: A Longitudinal Dataset and Benchmark for Dynamic Order Cancellation Prediction in On-Demand Ride-Sharing Systems

NeurIPS 2026 Evaluations & Datasets Track

Authors: Huayang Liu^1,, Mingyu Zheng^1,, Wei Zhang^1,2,†, Yijun Bian³, Kai Wang^1,2
1 School of Vehicle and Mobility, Tsinghua University
² State Key Laboratory of Intelligent Green Vehicle and Mobility, Tsinghua University
³ Department of Computer Science, University of Copenhagen
^* Equal contribution    ^† Corresponding author (weizhanghk@tsinghua.edu.cn)

• Project Page: https://tsingtransit.github.io/survcancel/
• Benchmark Code Repository: https://github.com/tsingtransit/survcancel

---

Overview

SurvCancel is a large-scale real-world longitudinal dataset and benchmark for dynamic passenger-order cancellation prediction in on-demand ride-sharing systems. Constructed from operational logs across four service regions (R1–R4), the release contains 173,037 released orders and 7,944,171 pre-pickup snapshots (~40.56 GiB uncompressed).

Each JSONL record represents a single focal order with its static trip attributes, terminal outcome (pre_match_cancel, post_match_cancel, or completed), and a time-ordered sequence of pre-pickup snapshots sampled on a 10-second grid. Every snapshot captures the observable operational context at that timestamp: the focal order state, surrounding active orders, available vehicles, and the assigned vehicle once matched.

Note that the four regional files (train-R1.jsonl through train-R4.jsonl) use the train-* prefix to follow Hugging Face's default file-sharding convention; actual order-level train/validation/test splits (in-distribution, leave-one-region-out, and chronological) are deterministically constructed by the official benchmark codebase.

Release Statistics

Region	Active Days	Orders	Snapshots	Pre-Match Cancel	Matched	Post-Match Cancel	Completed
R1	38	60,548	2,717,147	26,511	34,037	10,987	23,050
R2	72	6,518	240,075	310	6,208	872	5,336
R3	72	42,106	2,099,571	13,404	28,702	7,667	21,035
R4	72	63,865	2,887,378	10,449	53,416	10,651	42,765
Total	—	173,037	7,944,171	50,674	122,363	30,177	92,186

Repository Files

• train-R1.jsonl, train-R2.jsonl, train-R3.jsonl, train-R4.jsonl: Full order-level sequence shards for regions R1–R4 (43,550,501,224 bytes total).
• sample/sample_records.jsonl: Small subset of records for quick inspection and smoke testing.
• SCHEMA.md: Field-level specification for orders, snapshots, active orders, and available vehicles.

Loading the Dataset

Option 1: Stream with Hugging Face datasets

Stream all four regions:

from datasets import load_dataset

dataset = load_dataset(
    "survcancel-anonymous/SurvCancel",
    split="train",
    streaming=True,
)

Stream a single region shard directly:

from datasets import load_dataset

r1 = load_dataset(
    "json",
    data_files="hf://datasets/survcancel-anonymous/SurvCancel/train-R1.jsonl",
    split="train",
    streaming=True,
)

Option 2: Download and Build the Clean58 Benchmark Tensors

To reproduce the paper's benchmark protocols (in-distribution, leave-one-region-out, and chronological splits) and construct the standardized 58-dimensional Clean58 arrays, use the official GitHub repository:

git clone https://github.com/tsingtransit/survcancel.git
cd survcancel
pip install -r requirements.txt

# 1. Download raw JSONL shards (resumable, SHA-256 verified)
bash scripts/download_data.sh ./raw_jsonl

# 2. Construct Clean58 arrays and fixed evaluation landmarks
bash scripts/prepare_clean58.sh ./raw_jsonl ./data/clean58

De-Identification and Privacy

To protect operational and passenger privacy under the data-sharing agreement:

• Order, vehicle, and region identifiers are replaced with 16-character salted HMAC-SHA256 surrogates (ord_*, veh_*, reg_*). Within each namespace, identical source entities map to the same surrogate so longitudinal trajectories and vehicle assignments remain linkable.
• Spatial coordinates (pickup_x, pickup_y, dropoff_x, dropoff_y, vehicle_x, vehicle_y) are transformed into standardized planar coordinates rather than raw GPS latitude/longitude.
• Raw station identifiers, cryptographic salts, and identifier mapping tables are excluded from the release.

Intended Use and License

This dataset is intended solely for academic and non-commercial research on dynamic survival analysis, time-to-event modeling, and on-demand urban mobility operations.

• Dataset License: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0).
• Benchmark Code License: MIT License (tsingtransit/survcancel).

Citation

If you use the SurvCancel dataset or benchmark in your research, please cite:

@inproceedings{liu2026survcancel,
  title     = {SurvCancel: A Longitudinal Dataset and Benchmark for Dynamic Order Cancellation Prediction in On-Demand Ride-Sharing Systems},
  author    = {Liu, Huayang and Zheng, Mingyu and Zhang, Wei and Bian, Yijun and Wang, Kai},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Track on Evaluations and Datasets},
  year      = {2026},
  url       = {https://tsingtransit.github.io/survcancel/}
}