HHAR: Heterogeneity Human Activity Recognition Dataset

Overview

HHAR investigates variation in motion sensing across device models. The activity experiment involves nine users, eight smartphones and four smartwatches. Its six named activities are biking, sitting, standing, walking, stairs up and stairs down.

Verified raw activity files

Total: 33,741,500 rows. Each of the four files has ten columns: Index, Arrival_Time, Creation_Time, x, y, z, User, Model, Device and gt. The ten-column schema applies to these activity files only.

Time, devices and labels

Arrival_Time records arrival at the collecting application; Creation_Time is the timestamp attached by the operating system. The README directs users to the paper for timestamp semantics. Sensors were sampled at the highest rate each device allowed, not a common fixed frequency.

The phone accelerometer file contains eight device identifiers, but the phone gyroscope file contains six; samsungold_1 and samsungold_2 are absent from the latter. All four activity files contain participant identifiers a to i and the six named activities plus the literal null class. Treat null labels separately from numeric missingness.

Still experiment and count discrepancy

The separate still archive contains 198 measurement CSVs with 2,608,312 data rows in total. Their schemas vary: 163 files use Creation_Time, Sensor_Time, Arrival_Time, x, y, z; 34 use Index, Arrival_Time, Creation_Time, x, y, z; one uses Arrival_Time, Creation_Time, x, y, z. Six x-sense.csv files additionally have a trailing comma after each data row, creating a seventh empty field despite a six-column header.

The catalog row and column counts, 33,741,500 rows and 10 columns, describe the four activity files, which share one schema. The still archive's 2,608,312 rows are reported separately because its schemas differ; the two experiments together contain 36,349,812 CSV data rows. Neither this total nor the activity-only total reproduces UCI's headline of 43,930,257 instances. The reason for that discrepancy was not established.

Source and license

Source: UCI Heterogeneity Activity Recognition, donated 25 October 2015. Dataset DOI: 10.24432/C5689X. The source citation names Henrik Blunck, Sourav Bhattacharya, Thor Prentow, Mikkel Kjærgaard and Anind Dey.

Companion paper: Smart Devices are Different: Assessing and Mitigating Mobile Sensing Heterogeneities for Activity Recognition (2015). The paper and the dataset have different author lists.

The UCI record licenses the data under CC BY 4.0. Reuse requires appropriate attribution, a license link and an indication of changes. This catalog description is an independently prepared summary.