FitRec Datasets

These datasets contain user sport records from Endomondo. Data includes multiple sources of sequential sensor data such as heart rate, speed, GPS as well as sport type, user gender and weather condition (i.e. temperature, humidity).

We collected these datasets for academic use only. Please do not redistribute them or use for commercial purposes.

If you are using our dataset, please cite the following papers:

  • Jianmo Ni, Larry Muhlstein, Julian McAuley, "Modeling heart rate and activity data for personalized fitness recommendation", in Proc. of the 2019 World Wide Web Conference (WWW'19), San Francisco, US, May. 2019.

If you find any question or problem with the data, please contact Jianmo Ni (jin018@ucsd.edu). Your help would be appreciated.

In this work, we have prepared three versions of the dataset:

The raw data:

  • 253,020 workouts / 1,104 users

This is the original dataset without cleaning. It contains meta data that we have not studied in this work such as weather condition. Feel free to play with it.

Following is the download link:

The filtered version for workout route prediction task:

  • 167,373 workouts / 956 users

We use heuristics to clean the data by filtering out those abnormal workout samples such as overly large magnitude, mismatching timestamps, abrupt changes in GPS coordinates. We also derive multiple variables, e.g., speed and distance from the measurements.

After getting the filter dataset (i.e. endomondoHR_proper.json), we process the dataset by normalizing the measurement data into Z-Scores (i.e. that's why the sample showing below includes both positive and negative values). During test, we further filter out those users with less than 10 workouts.

During prediction, we try to predict either tar_heart_rate or tar_derived_speed, which are the heart rate and derived_speed under the original scale (i.e. before normalization).

Note: since_began and since_last mean are the number of seconds compared with the current workout and the first/most recent workout of that user. We don't use that in our work.

Followings are the download link:

Example of one workout sample from endomondoHR_proper.json:

userId: 10921915
gender: male
sport: bike
id: 396826535
longitude: [24.64977040886879, 24.65014273300767, 24.650910682976246, 24.650668865069747, 24.649145286530256, ...]
latitude: [60.173348765820265, 60.173239801079035, 60.17298021353781, 60.172477969899774, 60.17186114564538, ...]
altitude: [-1.8044666444624418, -1.8190453555595787, -1.8190453555595787, -1.8511185199732794, -1.871528715509271, ...]
timestamp: [1408898746, 1408898754, 1408898765, 1408898778, 1408898794, ...]
time_elapsed: [-0.12256752559145224, -0.12221090169596584, -0.12172054383967204, -0.12114103000950663, -0.12042778221853381, ...]
heart_rate: [-8.197369036801112, -5.867841701016304, -3.961864789919643, -4.173640002263717, -3.961864789919643, ...]
derived_speed: [-7.0829444390064396, -2.8061928357004815, -0.3976286593020398, -0.7571073884764162, 2.6415189187026646, ...]
distance: [-4.372303649217691, -2.374952819539426, -0.07926348591212737, 0.4284751220389811, 4.710835498111755, ...]
tar_heart_rate: [100, 111, 120, 119, 120, ...]
tar_derived_speed: [0, 10.751376415573548, 16.806294372816662, 15.902596545765366, 24.446443398153843, ...]
since_begin: [1378478.8892184314, 1378478.8892184314, 1378478.8892184314, 1378478.8892184314, 1378478.8892184314, ...]
since_last: [2158.84607810351, 2158.84607810351, 2158.84607810351, 2158.84607810351, 2158.84607810351, ...]

The re-sampled version for short term heart rate prediction task

  • 102,343 workouts / 887 users

Based on the processed dataset above (the normalized version), we use interpolation to obtain a dataset with same sampling intervals (i.e. 10 seconds) during each workout. To obatain a valid interpolation, we further clean the data once via timestampTest and filtered out abnormal samples after interpolation. You can check more details in our paper.

Following is the download link for the data, we provide both the numpy format and JSON format:

Loading data:

To load the npy file, you can use:

path = Path("data/")
out_path = str(path / "processed_endomondoHR_proper.npy")
data = np.load(out_path)[0]

To load the json file, you can use:

data = []
#with gzip.open('endomondoHR.json.gz') as f:
with open('endomondoHR_proper.json') as f:
    for l in f:
        data.append(eval(l))