Quantifying Smartphone Overuse: A Machine Learning Analysis of Severity and Lifestyle Effects

Abstract

Smartphone use has become an important part of communication, education, entertainment and social interaction. At the same time, excessive or uncontrolled usage may be associated with reduced productivity, disturbed routines, social difficulties and other lifestyle concerns. This paper presents a machine-learning-based approach for identifying patterns of problematic smartphone usage (PSU) from demographic and smartphone-behavior responses. The work is based on the project implementation documented in the accompanying final report. The dataset contains 501 responses and 21 columns, including behavioral questions related to checking the phone, use during social situations, use before sleep, gaming, notifications and other usage habits. The original target distribution contains 322 non-addicted and 179 addicted responses. Random over-sampling is used in the documented implementation to create a balanced dataset of 644 records. The processing pipeline includes data upload, preprocessing, categorical conversion, and class balancing, model training, prediction and performance analysis. Three implemented classifiers are examined: a Stacking Classifier using Random Forest, SVC, XGBoost and K-Nearest Neighbors as base learners with Logistic Regression as the meta-model; CatBoost; and Extra Trees. The reported test results are approximately 95% accuracy for Stacking, 96% for CatBoost and 100% for Extra Trees on the documented test runs. The paper presents the system architecture, input/output interface, model evaluation, confusion matrices and practical implications. The reported results demonstrate the feasibility of a web-based machine-learning system for behavioral pattern classification, while larger and independently collected datasets are required for stronger generalization and real-world validation.