seahorse-stpp/citibike-stpp

CitiBike STPP Benchmark Dataset

A benchmark-ready Spatio-Temporal Point Process (STPP) dataset derived from CitiBike NYC trip data (2019), following the official split semantics of the Neural STPP paper.

---

Dataset Description

Each record represents a sequence of bike-trip departure events associated with a specific start station on a specific day in 2019.
The dataset covers April–August 2019, partitioned into train / val / test by date range.

---

Source Format

Raw data was stored in a NumPy .npz archive (citibike_2019.npz).
Each key in the archive is a string of the form YYYYMMDD_XXX, where:

• YYYYMMDD is the date (e.g. 20190401)
• XXX is a zero-padded station index (e.g. 000–019)

Each key maps to a (N, 3) float64 array with columns [t, x, y].

---

Sequence Unit

One .npz key = one sequence.
No new windowing or segmentation was applied. The original benchmark sequence unit from the Neural STPP repo is preserved exactly.

---

Event Schema

Field	Type	Description
t	float	Time of event — fractional hours within the day (e.g. 1.06 = 1:03 AM)
x	float	Longitude of trip start station (NYC range ≈ −74.02 to −73.91)
y	float	Latitude of trip start station (NYC range ≈ 40.68 to 40.80)

Values are exported as-is — no normalization applied.
The Neural STPP codebase applies StdScaler normalization at training time, not during preprocessing.

---

Split Semantics

Split	Date Range	Keys matching	Sequences
train	2019-04-01–2019-07-31	20190[4567]DD_XXX	2440
val	2019-08-01–2019-08-15	201908DD_XXX where DD ≤ 15	300
test	2019-08-16–2019-08-31	201908DD_XXX where DD > 15	320

Split logic mirrors CitibikeDS.splits in the Neural STPP codebase — no random splitting, no reshuffling.

---

File Structure

citibike-stpp/
├── train.jsonl        # 2440 sequences
├── val.jsonl          # 300 sequences
├── test.jsonl         # 320 sequences
├── splits.json        # {"train": [...], "val": [...], "test": [...]}
├── dataset_meta.json  # Task/schema metadata
└── README.md

---

JSONL Row Schema

Each line in a .jsonl file is a JSON object:

{
  "sequence_id": "20190401_000",
  "events": [
    {"t": 1.062, "x": -73.976, "y": 40.751},
    {"t": 2.318, "x": -73.991, "y": 40.748},
    ...
  ]
}

Example (Python)

import json

with open("train.jsonl") as f:
    for line in f:
        seq = json.loads(line)
        sid    = seq["sequence_id"]   # e.g. "20190401_000"
        events = seq["events"]        # list of {"t", "x", "y"} dicts
        t = [e["t"] for e in events]
        x = [e["x"] for e in events]
        y = [e["y"] for e in events]

---

Source & License

• Source data: CitiBike System Data
  Made available under the CitiBike Data License Agreement.
• Processing code reference: facebookresearch/neural_stpp (MIT License)

Version 1.0.1 Time Ordering Fix

The original 1.0.0 export preserved source event order in train.jsonl, but 255 train sequences contained non-increasing adjacent event times. Version 1.0.1 applies a deterministic data-level repair: events are stable-sorted by t within each sequence, and the single remaining exact timestamp tie is advanced with numpy.nextafter(prev_t, +inf). Validation requires np.diff(times) > 0 for every sequence in train/val/test.