Interactive Illustration for Sparse Step Counts Data Inference
Step counts data is a measurement of physical activity, and can be collected expediently with
smartphones and wearable devices. Complete high time-resolution step counts data records
the time and intensity of physical activity, and reflect an individual's daily
routine. Mining activity patterns that reflect such daily routine is the foundation of many
applications such as human behavior modeling, behavior prediction and intervention,
and pervasive healthcare.
In practice, however, obtaining complete step counts data may be difficult and costly.
Storage and power constraints may limit the number of data points acquired at the hardware level.
Usage and data permission may also limit the number of data points available at the application
level. Such limitations will result in step counts data with varied sparsity. Inferring activity
patterns from sparse step counts data can not only find its value in understanding user’s behavior,
it also may provide useful insights into the design of cost effective hardware and software.
In paper "Inferring Activity Patterns from Sparse Step Counts Data with Recurrent Neural Networks", we design a deep learning model implementing data-driven imputation and classification in an end-to-end manner for sparse step counts data. This webpage is developed to demonstrates our study. It has two functions:
- Generation and inference of one sparse step counts data sample.
- Activity pattern inference of sparse step counts dataset.
You can download data and codes subject to our terms of use.
Inference of one sparse step counts sample
- The generation of sparse step counts sample.
- Choosing granularity of activity pattern
TIPS:the smaller the input value is, the more sparse the sample is.
Fine granular activity patters with 70 categories have 5 quantization levels in intensity and 4 time windows which is illustrated in the following picture. The interval between two adjacent intensity levels is 0.25 (means one level of intensity is 25% of total step counts). Keep the number of time windows unchanged, medium granular activity patterns with 15 categories have 3 levels in intensity and coarse granular activity patterns with 5 categories have 2 levels in intensity. And the interval between two adjacent intensity levels of medium granularity and coarse granularity are 0.5 and 1.0, respectively.
![](/static/picture1.png)
The illustration of fine granularity activity patterns. The classes' centers are visualized by the 70 stepwise curves (in red) from (0, 0.00) to (24, 1.00). The red dotted line is the centroid of the class that the sample belongs to and it represents a pattern of (0.25, 0.75, 0.75, 1), which means 25% of total steps are taken around 8 AM, 50% around noon, and the remaining 25% around 8 PM.
Inference of sparse step counts dataset
We also support infer sparse step counts data in batches. You can upload the dataset and we will return the result in this website. But there are some requirements for the format of uploaded file.
- We only support csv file.
- The number of columns in csv file must be 290. The first column is total steps. The second column is the number of data points. And the last 288 columns are cumulative step counts data with intervals of 5 minutes throughout the day. If there is no data in some minutes, the value of corresponding windows will be set to null.
- The number of data points shouldn't be 0. And the cumulative step counts data cannot be all null.
The following table is an example.
Total Steps | # Data Points | 0 | 1 | 2 | 3 | ... | 130 | 131 | 132 | 133 | ... | 285 | 286 | 287 |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
8200 | 5 | ... | 4500 | ... | 8200 | |||||||||
12250 | 8 | 0 | ... | 3000 | 3500 | 3781 | ... | 12250 | 12250 |
In order to get the correct result, please make sure that the uploaded file meets the above requirements.