README.md

---
tags:
- time series
- time series classification
- monster
- HAR
license: other
pretty_name: WISDM2
size_categories:
- 100K<n<1M
---
Part of MONSTER: <https://arxiv.org/abs/2502.15122>.

|WISDM2||
|-|-:|
|Category|HAR|
|Num. Examples|149,034|
|Num. Channels|3|
|Length|100|
|Sampling Freq.|20 Hz|
|Num. Classes|6|
|License|Other|
|Citations|[1]|

***WISDM2*** extends the original *WISDM* dataset by collecting data in real-world environments using the Actitracker system.  This system was designed for public use and provides a more extensive collection of sensor readings from users performing the same six activities.  The dataset contains 2,980,765 samples with three dimensions, and the data was recorded from a larger and more diverse set of participants in naturalistic settings, offering a valuable resource for real-world activity recognition [1]. The processed dataset has 149,034 time series, each with length 100 (again, representing 5 seconds of data at a sampling rate of 20 Hz).  WISDM2 is split based on subjects.

[1] Gary Mitchell Weiss and Jeffrey Lockhart. (2012). The impact of personalization on smartphone-based activity recognition. In Workshops at the *26<sup>th</sup> AAAI Conference on Artificial Intelligence*.