Feature Engineering

This repository contains the code to generate features from the raw data available on the LivePeople data catalog powered by Datascientia. The following code has been tested on DivesityOne dataset.

All features in this repository are extracted over fixed-length time intervals.

Each user's data is divided into consecutive, non-overlapping time windows (e.g., 30-minute segments), and features are computed independently for each interval. This time-based aggregation applies to all supported sensor types (e.g., accelerometer, screen, wifi, etc.).

Typical usage of the features is for training machine learning models and for people’s everyday life behavior analysis with mobile data, e.g., behavioral changes over time.

Details about the datasets can be found on LivePeople data catalog.

A sample test dataset is provided in the test/ directory for quick experimentation and development.

Requirements

• Python 3.9
• Snakemake (for workflow automation)

Installation

Clone the repository

bash git clone git@github.com:datascientiafoundation/feature-engineering.git

Create a new environment with Conda

bash conda env create -f environment.yml conda activate feature_env

To install Snakemake, use the following command:

bash pip install snakemake

Repository structure

``` . ├── config/ │ ├── config.yaml # Contains a list of countries and sensors ├── data/ │ ├── CREP/ # Datasets following the CREP structure (Hierarchical) │ ├── interim/ # Intermediate files, single sensor features │ ├── raw/ # Raw datasets │ └── processed/ # Processed datasets, aggregated features ├── logs/ ├── src/ │ ├── utils/ │ │ ├── appcategories.csv # Contains mapping related to application categories │ │ └── utils.py # Utility functions for the project │ ├── config.py # Handles loading or managing configuration settings │ ├── contribution.py # Logic related to computing or managing user contributions │ ├── feature.py # Script for performing feature engineering on a single dataset │ ├── join_features.py # Script for merging/joining features from different sensors │ └── load.py # Script for loading datasets ├── test/ # Test dataset (input, output) ├── CITATION.cff ├── environment.yml # Conda environment configuration file ├── LICENCE # Contains the license information for the project (e.g., MIT License) ├── README.md └── Snakefile # Defines the workflow for processing datasets

```

Data preparation

Download the dataset (the datasets can be requested on the Datascientia platform) and save it in the appropriate folder:

• If the dataset follows the flattened structure, save it in the data/raw folder.

• If the dataset follows the Hierarchical structure (as retrieved from the catalog), save it in the data/CREP folder. Datasets in this structure need to be flattened first by running the following script:

bash python -m src.load -i data/CREP -o data/raw -l logs/load.log

Workflow Description

The Snakemake workflow orchestrates the end-to-end processing of datasets located in the data/raw directory. The pipeline proceeds through the following stages:

1. Process Time Diary
   The workflow begins by processing the time diary data, which defines valid activity intervals for each user.

2. Process Single Sensors
   Using the time intervals extracted from the processed timediary, each sensor dataset is processed individually to extract features. The resulting features are stored in the data/interim directory.

3. Join Features
   Finally, all single-sensor features are joined based on their timestamps, resulting in a unified, time-aligned dataset saved in the data/processed directory.

To run the entire workflow, execute the following command: bash snakemake all --cores 1

Configuration

config.yaml includes the following settings:

• Available Datasets:
  A strict list of dataset names. Only the sensors explicitly listed here will be processed—any sensor not listed will be ignored, even if present.

• Time Diary Inclusion:
  A boolean flag that controls whether time diary data should be included in feature generation.

• Sensor Frequency (freq):
  Specifies the duration of each time window in minutes.

  • When using time diary:
    For instance, if freq is set to 30 minutes and time diary entries occur every 30 minutes, each time window spans from 15 minutes before to 15 minutes after a time diary timestamp. Sensor events within these windows are aggregated to generate features.
  • When not using time diary:
    The entire sensor data timeline is divided into consecutive, non-overlapping intervals of fixed length. For example, if freq is set to 30 minutes, intervals like [10:00–10:30), [10:30–11:00), and so on are created. Sensor events are assigned to these intervals based on their timestamps, and features are aggregated accordingly.

Process single dataset

You can process a single dataset using the src/feature.py script directly: bash python -m src.feature -i data/raw/<SENSOR>.parquet -t data/interim/timediary.csv -o data/interim/<SENSOR>.csv -l logs/<SENSOR>.log -f <FREQ> -ti True Example: bash python -m src.feature -i data/raw/accelerometer.parquet -t data/interim/timediary.csv -o data/interim/accelerometer.csv -l logs/accelerometer.log -f 30 -ti True

## Workflow without timediary if you want to process datasets without timediary invervals, you can change the timediary parameter in config/config.yaml to False. Then the run the workflow as: bash snakemake all --cores 1

To process single datset without timediary:

bash python -m src.feature -i data/raw/<SENSOR>.parquet -t data/interim/timediary.csv -o data/interim/<SENSOR>.csv -l logs/<SENSOR>.log -f <FREQ> -ti False

Cyclical Feature Encoding

Cyclical features, such as the hour of the day, are also added to the dataset in order to capture the circular nature of time. This encoding technique helps models understand that certain times are close to others, such as 23:00 being close to 00:00. By using sine and cosine transformations, these features are represented in a way that preserves their cyclical nature, which improves the performance of models that involve time-based patterns.

Source: Encoding Cyclical Features

Supported Datasets and Extracted Features

accelerometer

• x_min, x_max, x_mean, x_std
• y_min, y_max, y_mean, y_std
• z_min, z_max, z_mean, z_std
• magnitude_min, magnitude_max, magnitude_mean, magnitude_std

accelerometeruncalibrated

• x_min, x_max, x_mean, x_std
• y_min, y_max, y_mean, y_std
• z_min, z_max, z_mean, z_std
• xunc_min, xunc_max, xunc_mean, xunc_std
• yunc_min, yunc_max, yunc_mean, yunc_std
• zunc_min, zunc_max, zunc_mean, zunc_std
• magnitude_min, magnitude_max, magnitude_mean, magnitude_std

activities

• activity_Running ,activity_Unknown, activity_Tilting, activity_OnBicycle, activity_InVehicle, activity_Still, activity_Walking, activity_OnFoot,

airplanemode

• airplanemode_True, airplanemode_False

ambienttemperature

• min, std, min, max

applications

• app_category_nunique
• [application groups]
• app_nunique, app_entropy_basic

batterycharge

• battery_charging_ac, battery_no_charging, battery_charging_unknown

batterylevel

• battery_timestamp_first, battery_timestamp_last, battery_level_first, battery_level_last, battery_scale_mean, battery_delta

bluetooth

• bluetooth_addr_nuinque, bluetooth_mean, bluetooth_min, bluetooth_max, bluetooth_std, bluetooth_var, bluetooth_entropy_basic,

cellularnetwork

• cellular_lte_mean, cellular_lte_min, cellular_lte_max, cellular_lte_std,
• cellular_lte_entropy_basic, cellular_lte_num_of_devices

doze

• doze_True, doze_False

geomagneticrotationvector

• x_min, x_max, x_mean, x_std
• y_min, y_max, y_mean, y_std
• z_min, z_max, z_mean, z_std
• magnitude_min, magnitude_max, magnitude_mean, magnitude_std
• accuracy_min, accuracy_max, accuracy_mean, accuracy_std
• scalar_min, scalar_max, scalar_mean, scalar_std

gravity

• x_min, x_max, x_mean, x_std
• y_min, y_max, y_mean, y_std
• z_min, z_max, z_mean, z_std
• magnitude_min, magnitude_max, magnitude_mean, magnitude_std

gyroscope

• x_min, x_max, x_mean, x_std
• y_min, y_max, y_mean, y_std
• z_min, z_max, z_mean, z_std
• magnitude_min, magnitude_max, magnitude_mean, magnitude_std

gyroscopeuncalibrated

• x_min, x_max, x_mean, x_std
• y_min, y_max, y_mean, y_std
• z_min, z_max, z_mean, z_std
• xunc_min, xunc_max, xunc_mean, xunc_std
• yunc_min, yunc_max, yunc_mean, yunc_std
• zunc_min, zunc_max, zunc_mean, zunc_std
• magnitude_min, magnitude_max, magnitude_mean, magnitude_std

headsetplug

• headset_False, headset_True

light

• min, std, min, max

linearacceleration

• x_min, x_max, x_mean, x_std
• y_min, y_max, y_mean, y_std
• z_min, z_max, z_mean, z_std
• magnitude_min, magnitude_max, magnitude_mean, magnitude_std

location

• latitude, longitude,
• longitude_mean, longitude_min, longitude_max,
• latitude_mean, latitude_min, latitude_max,
• altitude_mean, altitude_min, altitude_max,
• speed_mean, speed_min, speed_max, speed_std,
• radius_of_gyration, distance_sum

magneticfield

• x_min, x_max, x_mean, x_std
• y_min, y_max, y_mean, y_std
• z_min, z_max, z_mean, z_std
• magnitude_min, magnitude_max, magnitude_mean, magnitude_std

magneticfielduncalibrated

• x_min, x_max, x_mean, x_std
• y_min, y_max, y_mean, y_std
• z_min, z_max, z_mean, z_std
• xunc_min, xunc_max, xunc_mean, xunc_std
• yunc_min, yunc_max, yunc_mean, yunc_std
• zunc_min, zunc_max, zunc_mean, zunc_std
• magnitude_min, magnitude_max, magnitude_mean, magnitude_std

music

• music_False, music_True

notification

• notification_posted, notification_removed

orientation

• x_min, x_max, x_mean, x_std
• y_min, y_max, y_mean, y_std
• z_min, z_max, z_mean, z_std
• magnitude_min, magnitude_max, magnitude_mean, magnitude_std

pressure

• min, std, min, max

proximity

• min, std, min, max

relativehumidity

• min, std, min, max

ringmode

• ringmode_mode_silent, ringmode_mode_normal, ringmode_mode_vibrate

rotationvector

• x_min, x_max, x_mean, x_std
• y_min, y_max, y_mean, y_std
• z_min, z_max, z_mean, z_std
• magnitude_min, magnitude_max, magnitude_mean, magnitude_std
• accuracy_min, accuracy_max, accuracy_mean, accuracy_std
• scalar_min, scalar_max, scalar_mean, scalar_std

screen

• screen_SCREEN_ON, screen_SCREEN_OFF, screen_episodes_count,
• screen_mean_seconds_per_episode, screen_min_seconds_per_episode, screen_max_seconds_per_episode, screen_std_seconds_per_episode

stepcounter

• steps_counter

stepdetector

• steps_detected_count

touch

• touch_count

wifi

• is_connected

wifinetworks

• wifi_num_of_devices,
• wifi_mean_rssi, wifi_min_rssi, wifi_max_rssi, wifi_std_rssi