bigchestnut/nyc_slot_daily_trajectories

Dataset Card for NYC Daily Trajectories in 2hour time slot

Dataset Summary

nyc_daily_trajectories_sae.json contains 65,115 fixed-length daily trajectories derived from the TSMC2014 NYC Foursquare check-ins. Each record aggregates one user-day of check-ins at a 2-hour granularity (00:00–22:00) so the data is immediately consumable by sparse autoencoders and other sequential encoders that expect constant-width vectors. Locations are rounded to 0.001 degrees and missing slots are filled via last-observation-carried-forward with an optional user-specific home fallback, ensuring there are no unknown placeholders.

Supported Tasks and Leaderboards

• —representation-learning: learn interpretable latent structure of daily activity routines.
• —trajectory-completion: treat the sequence as conditional inputs for forecasting the next slot or inferring missing slots.
• —persona-clustering: cluster daily routines using embeddings for downstream personalization experiments.

Languages

English (activity names originate from the original Foursquare metadata).

Dataset Structure

Data Instances

{
  "user_id": "69",
  "date": "2012-04-03",
  "day_of_week": "Tuesday",
  "00:00_lat": 40.716,
  "00:00_lon": -73.883,
  "00:00_activity": "Home (private)",
  ...,
  "22:00_lat": 40.713,
  "22:00_lon": -73.879,
  "22:00_activity": "Bar"
}

Data Fields

• —user_id: anonymized user identifier string from TSMC2014.
• —date: local date (YYYY-MM-DD).
• —day_of_week: weekday name derived from the local date.
• —For each slot HH:MM in {00:00, 02:00, …, 22:00}:
• —HH:MM_lat: rounded latitude (float, 3 decimal places).
• —HH:MM_lon: rounded longitude (float, 3 decimal places).
• —HH:MM_activity: activity or POI category string from Foursquare.

Data Splits

Single combined split; add train/val/test logic downstream if needed.

Data Collection

Source Data

• —TSMC2014 NYC Foursquare check-ins (Gowalla/Foursquare crawl, April 2012 timeframe).
• —Local timestamps reconstructed using the provided timezone offsets.

Annotations

No additional annotations beyond the original Foursquare category labels.

Considerations for Using the Data

Social Impact of Dataset

The data can surface how individuals spend their days in NYC and should be used responsibly, especially when building models that might influence mobility services.

Discussion of Biases

The underlying TSMC2014 dataset is biased toward Foursquare power users and over-represents certain urban demographics, neighborhoods, and venue types. Treat learned patterns as reflective of that community rather than the entire NYC population.

Known Limitations

• —Only users with at least one check-in (or a discovered home) per day are retained, so very sparse user-days are removed.
• —Activities are limited to Foursquare category strings; no semantic normalization beyond lowercasing for home detection.
• —Spatial precision is coarse (≈100 m) due to rounding, by design for privacy and model stability.

Preprocessing and Augmentation

• —All coordinates rounded to 3 decimals.
• —Duplicate check-ins within a slot are dropped (keep earliest).
• —Missing slots seeded from prior observations; if no prior slot exists, the user’s inferred home (first venue with "home" in the activity) is used. User-days without either are omitted.