ABSTRACT

Smartphone battery life is a critical component of user satisfaction, especially in an

age where people are increasingly reliant on mobile devices. Understanding and
predicting battery drain based on user behavior and device characteristics can
enable both users and developers to optimize performance and energy efficiency.
This project proposes a machine learning-based solution to accurately estimate
daily battery drain (in mAh) using real-world usage data.

The dataset includes features such as app usage time, screen-on time, number of
apps installed, daily data usage, device model, operating system, age, and gender.
After performing data cleaning, feature encoding, and scaling, several regression
models were trained, including Linear Regression, Ridge Regression, Random
Forest, and Gradient Boosting. These models were evaluated using key
performance metrics like Mean Absolute Error (MAE), Root Mean Squared Error
(RMSE), and the R² score to measure the goodness of fit.

Among all models, the Random Forest Regressor demonstrated the highest
performance with an R² score of 0.95, indicating a strong ability to capture the
nonlinear relationships in the data. Hyperparameter tuning using
RandomizedSearchCV further optimized the model, improving its generalization
capability.

To make the solution user-friendly, an interactive interface was developed using

Jupyter Notebook widgets, allowing users to input their usage behavior and receive
real-time predictions for battery consumption. This not only makes the model
accessible but also showcases its practical utility in everyday scenarios.

Overall, the project bridges the gap between data science and real-world mobile
experiences, offering a predictive tool that could be integrated into future
smartphone systems or battery optimization apps.

1
 INTRODUCTION
In recent years, the widespread adoption of smartphones has transformed how
people communicate, work, and interact with the digital world. Mobile devices are
no longer just tools for calling or texting—they have evolved into multi-functional
platforms used for social networking, entertainment, education, productivity, and
much more. As users engage with a wide variety of apps daily, massive amounts of
behavioral data are generated, offering deep insights into user habits, preferences,
and routines.

Background

As smartphones become increasingly central to our daily lives—used for

communication, entertainment, productivity, and social media—the demand for
longer battery life grows exponentially. Despite advances in battery technology,
users still frequently experience rapid battery drain, often without a clear
understanding of the cause. Predictive modeling using machine learning offers a
promising solution by analyzing user behavior and device characteristics to
forecast battery usage and optimize energy consumption.

Problem Statement

Users often struggle to understand which factors contribute most to their phone’s
battery drainage. Traditional battery monitoring tools provide surface-level
statistics but fail to deliver personalized insights or predictions. There is a lack of
intelligent systems that can proactively inform users about their expected battery
consumption based on their daily phone usage patterns.

Objective

The primary objective of this project is to develop a machine learning-based model

that predicts the daily battery drain (measured in mAh) of a smartphone based on
various user and device-level features. Additionally, the project aims to identify the
most influential factors contributing to battery drain and provide a user-friendly,
interactive system to test different usage patterns and receive predictions.

2
Scope

This project covers:

 Preprocessing a dataset with features such as app usage time, screen-on

duration, number of apps, and data usage.

 Applying and comparing multiple regression models (Linear Regression,

XGBoost, Random Forest, and Gradient Boosting).

 Hyperparameter tuning using RandomizedSearchCV for optimization.

 Evaluating model performance using MAE, RMSE, and R².

 Building an interactive prediction interface for end users via Jupyter

Notebook.

The final model can be extended in future to support real-time integration with
mobile operating systems or apps focused on battery health and device
performance.

3
 DATASET OVERVIEW
The Mobile Device Usage and User Behavior dataset is designed to capture how
individuals interact with mobile applications in different contexts. It includes
detailed records of app usage sessions, offering valuable insights into digital
behavior such as time spent on apps, types of apps used, battery levels during use,
and time-based patterns.

This data can be used for behavior analysis, usage prediction, personalization,
and device optimization using machine learning models.

Link: [Link]
user-behavior-dataset

This dataset provides a comprehensive analysis of mobile device usage patterns

and user behavior classification. It contains 700 samples of user data, including
metrics such as app usage time, screen-on time, battery drain, and data
consumption. Each entry is categorized into one of five user behavior classes,
ranging from light to extreme usage, allowing for insightful analysis and modeling.

Tools and Technology Used:

Python (for modeling and preprocessing)

Libraries: pandas, numpy, scikit-learn, seaborn, matplotlib

Jupyter Notebook : For code development and experimentation

Machine Learning Models :

1. Linear Regression – a simple, interpretable model that assumes a linear

relationship between features and target.
2. Random Forest Regressor – an ensemble method that builds multiple
decision trees for better accuracy and robustness.
3. Gradient Boosting – a boosting technique that builds models sequentially to
reduce error.

4
 4. XGBoost – an advanced version of gradient boosting optimized for speed
and performance.

Dataset Content :

Column Name Description

User ID Unique identifier for each user. Helps differentiate data

across users.
Device Model Model of the user's smartphone (e.g., iPhone 13,
Samsung Galaxy S21).
Operating System The operating system of the device: either iOS or
Android.
App Usage Time Daily time spent on mobile applications, measured in
minutes.
Screen On Time Average number of hours per day the device screen is
active.
Battery Drain Amount of battery consumed daily, measured in mAh
(milliamp-hours).
Number of Apps Total number of apps currently installed on the device.
Installed
Data Usage Daily mobile data consumption in megabytes (MB).

5
 METHODOLOGY AND IMPLEMENTATION
System Architecture
The system is designed to predict battery drain (in mAh/day) based on user
behavior and device characteristics using machine learning regression models. It
consists of the following stages:
1. Data Preprocessing:
o Handling missing values and encoding categorical variables (e.g.,
Gender, Operating System).
o Feature scaling using StandardScaler to normalize numerical data.
2. Exploratory Data Analysis (EDA):
o Understanding the relationship between features and battery drain
using visualizations such as heatmaps, pair plots, and boxplots.
o Identifying correlations between features to aid in model selection and
interpretation.
3. Model Training:
o Applying multiple regression algorithms: Linear Regression, Ridge
Regression, Random Forest Regressor, and Gradient Boosting
Regressor.
o Evaluating each model’s performance using metrics like MAE,
RMSE, and R².
4. Hyperparameter Tuning:
o Using RandomizedSearchCV to tune parameters of models (especially
Random Forest and Gradient Boosting) for optimal performance.
o Performing cross-validation to ensure generalization and avoid
overfitting.
5. Interactive Prediction Interface:

6
 o A Jupyter Notebook interface is developed where users can input their
usage details and receive real-time predictions for battery drain.
Data Preprocessing
The dataset used includes features like screen time, app usage, number of apps,
data usage, and demographic info (age, gender). The preprocessing steps are:
1. Handling Missing and Categorical Data
Missing values are imputed (if any), and categorical variables (like Gender and
OS) are converted using one-hot encoding.

2. Feature Scaling
All numerical features are standardized using StandardScaler to ensure all models
(especially regularized ones) train effectively.

Exploratory Data Analysis (EDA)

Exploratory data analysis includes:

7
  Distribution Plot of Battery Drain:
A histogram with a KDE (Kernel Density Estimate) overlay is used to
visualize the distribution of battery drain values. This helps in understanding
the skewness and central tendency of the data. The plot indicates how
frequently certain battery drain values occur in the dataset.
 Feature Correlation with Battery Drain:
A horizontal bar plot is generated to show the correlation coefficients of all
numerical features with the target variable (Battery Drain). This helps in
identifying which features have a strong positive or negative relationship
with battery consumption, aiding in feature selection and model
interpretation.
 Boxplot by Operating System:
A boxplot is used to compare battery drain across different operating
systems (Android and iOS). This visualization highlights the distribution,
median, and presence of outliers in battery usage based on the user's device
OS.
 Boxplot by Gender:
Similarly, a boxplot is created to show battery drain trends based on gender.
It allows us to observe if there are any noticeable differences in battery
usage patterns between male and female users.

Model Training
Four regression models were implemented:
1. Linear Regression

8
 2. XGB Regressor
3. Random Forest Regressor
4. Gradient Boosting Regressor
Each model was trained using the preprocessed features, and evaluated using:
 MAE (Mean Absolute Error)
 RMSE (Root Mean Squared Error)
 R² Score (Coefficient of Determination)

Hyperparameter Tuning
Hyperparameter tuning was done using RandomizedSearchCV. Hyperparameter
tuning is a crucial step in improving model performance by finding the optimal set
of parameters. Instead of trying every possible combination (which can be
computationally expensive), RandomizedSearchCV randomly samples a specified
number of parameter combinations from a defined distribution. It uses cross-

9
validation to evaluate performance and select the best combination based on a
scoring metric

Interactive Prediction System

An interactive prediction interface was built using a Jupyter Notebook. It:
 Collects user input for features like app usage, screen-on time, age, and
gender.
 Preprocesses and scales the input.
 Uses the best trained model (Random Forest) to predict daily battery drain.

10
 RESULT
Model Evaluation Overview
After building and training multiple regression models to predict battery drain,
their performance was evaluated using standard regression metrics: Mean Absolute
Error (MAE), Root Mean Squared Error (RMSE), and R² Score. These metrics
provide insights into the accuracy and robustness of each model.
The following models were compared:
 Linear Regression

 XGB Regressor

 Random Forest Regressor

 Gradient Boosting Regressor

Performance Metrics
11
Model MAE (↓) RMSE (↓) R² Score (↑)
Linear Regression 160.32 mAh 195.83 0.938
mAh
XGB Regressor 163.51 mAh 201.85 0.934
mAh
Random Forest Regressor 145.40 mAh 171.82 0.952
mAh
Gradient Boosting 143.67 mAh 172.07 0.951
mAh

The Random Forest Regressor emerged as the best-performing model with the
lowest MAE and RMSE and the highest R² score. Its ensemble approach helped in
capturing complex patterns in the data effectively.
Visual Comparison of Models
The below bar plot visually compares the performance of all four models in terms
of MAE, RMSE, and R² score:

12
Hyperparameter Tuning (Randomized Search CV)
To further improve the model’s performance, RandomizedSearchCV was used
for hyperparameter optimization. It sampled multiple combinations of parameters
and evaluated their performance using cross-validation. This significantly
improved model generalization.

Interactive Battery Drain Prediction

The final tuned model was tested using a user-interactive input system, where users
could enter their own values such as screen-on time, data usage, and app usage,
and receive an estimated battery drain prediction.

13
14
 CONCLUSION
In this project, we developed a machine learning model aimed at predicting battery
drain on mobile devices based on various user behaviors and device features. By
leveraging regression models, including Random Forest, XGB Regressor, and
Gradient Boosting, we explored the relationships between features like app usage
time, screen-on time, data usage, and device specifications. The Random Forest
Regressor emerged as the best-performing model, with the lowest error metrics and
highest R² score, showcasing its ability to effectively capture the complex
dependencies of battery consumption.
Additionally, hyperparameter tuning using RandomizedSearchCV further
improved the model's accuracy, ensuring better generalization to new data. The
development of an interactive prediction tool also allowed users to input their data
and receive real-time predictions on battery drain, enhancing the practical value of
the model. Despite its success, the project highlighted areas for future
enhancement, including the inclusion of more diverse datasets and the exploration
of device-specific features for more precise predictions. Overall, this project
demonstrates the power of machine learning in mobile device optimization and
paves the way for future improvements in battery management systems.
Moreover, the project also provided insights into the importance of feature
engineering and data preprocessing in building a robust predictive model. The
careful selection and scaling of features, along with the application of
hyperparameter tuning, played a crucial role in enhancing model performance. The
visualizations, such as the distribution of battery drain and the feature correlation
heatmap, helped in understanding the underlying patterns in the data.

15
 BIBLIOGRAPHY
 Kaggle. (n.d.). Mobile device usage and user behavior dataset. Retrieved
from [Link]
and-user-behavior-dataset
 scikit-learn. (n.d.). scikit-learn documentation. Retrieved from [Link]
[Link]
 XGBoost. (n.d.). XGBoost documentation. Retrieved from
[Link]

16