rapids/docs/features/add-new-features.md

185 lines
11 KiB
Markdown

# Add New Features
!!! hint
We recommend reading the [Behavioral Features Introduction](/features/feature-introduction) before reading this page
!!! hint
You won't have to deal with time zones, dates, times, data cleaning or preprocessing. The data that RAPIDS pipes to your feature extraction code is ready to process.
## New Features for Existing Sensors
You can add new features to any existing sensors (see list below) by adding a new provider in three steps:
1. [Modify](#modify-the-configyaml-file) the `config.yaml` file
2. [Create](#create-a-provider-folder-script-and-function) a provider folder, script and function
3. [Implement](#implement-your-feature-extraction-code) your features extraction code
As a tutorial, we will add a new provider for `PHONE_ACCELEROMETER` called `VEGA` that extracts `feature1`, `feature2`, `feature3` in Python and that it requires a parameter from the user called `MY_PARAMETER`.
??? info "Existing Sensors"
An existing sensor is any of the phone or Fitbit sensors with a configuration entry in `config.yaml`:
- Phone Accelerometer
- Phone Activity Recognition
- Phone Applications Foreground
- Phone Battery
- Phone Bluetooth
- Phone Calls
- Phone Conversation
- Phone Light
- Phone Locations
- Phone Messages
- Phone Screen
- Phone WiFI Connected
- Phone WiFI Visible
### Modify the `config.yaml` file
In this step you need to add your provider configuration section under the relevant sensor in `config.yaml`. See our example for our tutorial's `VEGA` provider for `PHONE_ACCELEROMETER`:
??? example "Example configuration for a new accelerometer provider `VEGA`"
```yaml
PHONE_ACCELEROMETER:
TABLE: accelerometer
PROVIDERS:
RAPIDS:
COMPUTE: False
...
PANDA:
COMPUTE: False
...
VEGA:
COMPUTE: False
FEATURES: ["feature1", "feature2", "feature3"]
MY_PARAMTER: a_string
SRC_FOLDER: "vega"
SRC_LANGUAGE: "python"
```
| Key                          | Description
|---|---|
|`[COMPUTE]`| Flag to activate/deactivate your provider
|`[FEATURES]`| List of features your provider supports. Your provider code should only return the features on this list
|`[MY_PARAMTER]`| An arbitrary parameter that our example provider `VEGA` needs. This can be a boolean, integer, float, string or an array of any of such types.
|`[SRC_LANGUAGE]`| The programming language of your provider script, it can be `python` or `r`, in our example `python`
|`[SRC_FOLDER]`| The name of your provider in lower case, in our example `vega` (this will be the name of your folder in the next step)
### Create a provider folder, script and function
In this step you need to add a folder, script and function for your provider.
5. Create your provider **folder** under `src/feature/DEVICE_SENSOR/YOUR_PROVIDER`, in our example `src/feature/phone_accelerometer/vega` (same as `[SRC_FOLDER]` in the step above).
6. Create your provider **script** inside your provider folder, it can be a Python file called `main.py` or an R file called `main.R`.
7. Add your provider **function** in your provider script. The name of such function should be `[providername]_features`, in our example `vega_features`
!!! info "Python function"
```python
def [providername]_features(sensor_data_files, day_segment, provider, filter_data_by_segment, *args, **kwargs):
```
!!! info "R function"
```r
[providername]_features <- function(sensor_data, day_segment, provider)
```
### Implement your feature extraction code
The provider function that you created in the step above will receive the following parameters:
| Parameter&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; | Description
|---|---|
|`sensor_data_files`| Path to the CSV file containing the data of a single participant. This data has been cleaned and preprocessed. Your function will be automatically called for each participant in your study (in the `[PIDS]` array in `config.yaml`)
|`day_segment`| The label of the day segment that should be processed.
|`provider`| The parameters you configured for your provider in `config.yaml` will be available in this variable as a dictionary in Python or a list in R. In our example this dictionary contains `{MY_PARAMETER:"a_string"}`
|`filter_data_by_segment`| Python only. A function that you will use to filter your data. In R this function is already available in the environment.
|`*args`| Python only. Not used for now
|`**kwargs`| Python only. Not used for now
The code to extract your behavioral features should be implemented in your provider function and in general terms it will have three stages:
??? info "1. Read a participant's data by loading the CSV data stored in the file pointed by `sensor_data_files`"
``` python
acc_data = pd.read_csv(sensor_data_files["sensor_data"])
```
Note that phone's battery, screen, and activity recognition data is given as episodes instead of event rows (for example, start and end timestamps of the periods the phone screen was on)
??? info "2. Filter your data to process only those rows that belong to `day_segment`"
This step is only one line of code, but to undersand why we need it, keep reading.
```python
acc_data = filter_data_by_segment(acc_data, day_segment)
```
You should use the `filter_data_by_segment()` function to process and group those rows that belong to each of the [day segments RAPIDS could be configured with](/setup/configuration/#day-segments).
Let's understand the `filter_data_by_segment()` function with an example. A RAPIDS user can extract features on any arbitrary [day segment](/setup/configuration/#day-segments). A day segment is a period of time that has a label and one or more instances. For example, the user (or you) could have requested features on a daily, weekly, and week-end basis for `p01`. The labels are arbritrary and the instances depend on the days a participant was monitored for:
- the daily segment could be named `my_days` and if `p01` was monitored for 14 days, it would have 14 instances
- the weekly segment could be named `my_weeks` and if `p01` was monitored for 14 days, it would have 2 instances.
- the weekend segment could be named `my_weekends` and if `p01` was monitored for 14 days, it would have 2 instances.
For this example, RAPIDS will call your provider function three times for `p01`, once where `day_segment` is `my_days`, once where `day_segment` is `my_weeks` and once where `day_segment` is `my_weekends`. In this example not every row in `p01`'s data needs to take part in the feature computation for either segment **and** the rows need to be grouped differently.
Thus `filter_data_by_segment()` comes in handy, it will return a data frame that contains the rows that were logged during a day segment plus an extra column called `local_segment`. This new column will have as many unique values as day segment instances exist (14, 2, and 2 for our `p01`'s `my_days`, `my_weeks`, and `my_weekends` examples). After filtering, **you should group the data frame by this column and compute any desired features**, for example:
```python
acc_features["acc_rapids_maxmagnitude"] = acc_data.groupby(["local_segment"])["magnitude"].max()
```
The reason RAPIDS does not filter the participant's data set for you is because your code might need to compute something based on a participant's complete dataset before computing their features. For example, you might want to identify the number that called a participant the most throughout the study before computing a feature with the number of calls the participant received from this number.
??? info "3. Return a data frame with your features"
After filtering, grouping your data, and computing your features, your provider function should return a data frame that has:
- One row per day segment instance (e.g. 14 our `p01`'s `my_days` example)
- The `local_segment` column added by `filter_data_by_segment()`
- One column per feature. Your feature columns should be named `SENSOR_PROVIDER_FEATURE`, for example `accelerometr_vega_feature1`
??? example "`PHONE_ACCELEROMETER` Provider Example"
For your reference, this a short example of our own provider (`RAPIDS`) for `PHONE_ACCELEROMETER` that computes five acceleration features
```python
def rapids_features(sensor_data_files, day_segment, provider, filter_data_by_segment, *args, **kwargs):
acc_data = pd.read_csv(sensor_data_files["sensor_data"])
requested_features = provider["FEATURES"]
# name of the features this function can compute
base_features_names = ["maxmagnitude", "minmagnitude", "avgmagnitude", "medianmagnitude", "stdmagnitude"]
# the subset of requested features this function can compute
features_to_compute = list(set(requested_features) & set(base_features_names))
acc_features = pd.DataFrame(columns=["local_segment"] + ["acc_rapids_" + x for x in features_to_compute])
if not acc_data.empty:
acc_data = filter_data_by_segment(acc_data, day_segment)
if not acc_data.empty:
acc_features = pd.DataFrame()
# get magnitude related features: magnitude = sqrt(x^2+y^2+z^2)
magnitude = acc_data.apply(lambda row: np.sqrt(row["double_values_0"] ** 2 + row["double_values_1"] ** 2 + row["double_values_2"] ** 2), axis=1)
acc_data = acc_data.assign(magnitude = magnitude.values)
if "maxmagnitude" in features_to_compute:
acc_features["acc_rapids_maxmagnitude"] = acc_data.groupby(["local_segment"])["magnitude"].max()
if "minmagnitude" in features_to_compute:
acc_features["acc_rapids_minmagnitude"] = acc_data.groupby(["local_segment"])["magnitude"].min()
if "avgmagnitude" in features_to_compute:
acc_features["acc_rapids_avgmagnitude"] = acc_data.groupby(["local_segment"])["magnitude"].mean()
if "medianmagnitude" in features_to_compute:
acc_features["acc_rapids_medianmagnitude"] = acc_data.groupby(["local_segment"])["magnitude"].median()
if "stdmagnitude" in features_to_compute:
acc_features["acc_rapids_stdmagnitude"] = acc_data.groupby(["local_segment"])["magnitude"].std()
acc_features = acc_features.reset_index()
return acc_features
```
## New Features for Non-Existing Sensors
If you want to add features for a device or a sensor that we do not support at the moment (those that do not appear in the `"Existing Sensors"` list above), [contact us](/team) or request it on [Slack](/) and we can add the necessary code so you can follow the instructions above.