WISDM (Wireless Sensor Data Mining)

• Modality: Smartphone accelerometer + gyroscope, smartwatch accelerometer
• Primary Tasks: Human activity recognition, gesture recognition, sensor fusion
• Scale: 51 subjects, 18 activities, 20 Hz sampling (v1); 75 subjects and 7 activities in WISDM v2
• License: Creative Commons Attribution-NonCommercial-ShareAlike 3.0
• Access: https://www.cis.fordham.edu/wisdm/dataset.php

Summary

WISDM provides in-the-wild data captured from Android smartphones and smartwatches carried or worn by participants. It is widely used for accelerometer-based HAR baselines, lightweight models, and semi-supervised learning thanks to its accessible format (CSV) and straightforward activity labels.

Reference Paper

• Jennifer R. Kwapisz, Gary M. Weiss, Samuel A. Moore. "Activity Recognition using Cell Phone Accelerometers." SIGKDD Explorations, 2011. PDF

Benchmarks & Baselines

• Random Forest (feature-engineered) - Accuracy: 91.7%; Kwapisz et al., 2011.
• 1D CNN (DeepConvLSTM variant) - Accuracy: 95.8% (10-fold CV); widely cited reimplementations.
• Evaluation splits vary; ensure reproducibility by publishing subject IDs and random seeds.

Tooling & Ecosystem

• WISDM preprocessing scripts for windowing and labeling.
• TorchHAR includes loaders and baseline models.
• Scikit-multiflow supports incremental learning experiments with WISDM streams.

Known Challenges

• Sampling frequency varies slightly between devices; resample windows before training.
• Labels rely on participant compliance (self-report); expect occasional label noise.
• Older versions lack gyroscope data; cross-version comparisons must document sensor availability.

Cite

@article{kwapisz2011wisdm,
  title   = {Activity Recognition Using Cell Phone Accelerometers},
  author  = {Kwapisz, Jennifer R and Weiss, Gary M and Moore, Samuel A},
  journal = {SIGKDD Explorations},
  year    = {2011}
}