bigchestnut/tky_slot_daily_trajectories

Dataset Card for Tokyo Daily Trajectories SAE

Dataset Summary

tky_daily_trajectories_sae.json packages 67,575 daily trajectories distilled from the TSMC2014 Tokyo Foursquare check-ins. As with the NYC variant, every record covers a single user-day laid out on 12 two-hour slots with fixed lat/lon/activity triples for direct ingestion by sparse autoencoders or sequence models. UTC timestamps were converted to Japan-local time using the minute offsets included in the TSV source, and every slot is populated via carry-forward plus home fallback logic so there are no missing tokens.

Supported Tasks and Leaderboards

• representation-learning: learn latent routines specific to the Tokyo user population.
• mobility-simulation: drive agent-based or generative simulations using consistent per-slot features.
• anomaly-detection: identify unusual day patterns compared to habitual sequences.

Languages

English (original Foursquare venue category strings).

Dataset Structure

Data Instances

{
  "user_id": "1541",
  "date": "2012-04-04",
  "day_of_week": "Wednesday",
  "00:00_lat": 35.705,
  "00:00_lon": 139.620,
  "00:00_activity": "Cosmetics Shop",
  ...,
  "22:00_lat": 35.656,
  "22:00_lon": 139.734,
  "22:00_activity": "Housing Development"
}

Data Fields

Identical schema to the NYC release:

• user_id: anonymized TSMC2014 user identifier.
• date: local date string (YYYY-MM-DD).
• day_of_week: weekday name derived from the local date.
• Per-slot lat/lon/activity triples for 00:00, 02:00, …, 22:00 with 3-decimal rounding on coordinates.

Data Splits

Single combined split.

Data Collection

Source Data

• TSMC2014 Tokyo TSV file (dataset_TSMC2014_TKY.txt), parsed with the provided venue timezone offsets to obtain local timestamps.

Annotations

No additional annotations beyond the original venue category names.

Considerations for Using the Data

Social Impact of Dataset

Reveals stylized daily mobility routines of active Foursquare users in Tokyo; apply appropriate privacy and ethical safeguards when training downstream systems.

Discussion of Biases

The original dataset is biased toward heavy Foursquare users, popular neighborhoods, tourists, and English-labeled venues. Commercial areas and nightlife districts are likely over-represented relative to the broader population.

Known Limitations

• Sparse user-days without any observation or identifiable home were discarded, so the sample skews toward more active users.
• Some categories are inconsistent (e.g., duplicated spaces or punctuation) because they mirror the raw Foursquare metadata.
• Geographic rounding to 0.001 degrees (roughly 100 m) reduces spatial fidelity but promotes privacy and stable learning.

---

license: mit