Abstract:
The excessive use of social media by students has become a concern because it may hinder their
academic progress, mental health, and overall well-being. It examines how various machine
learning and deep learning models can predict and classify a student’s social media addiction
using demographic details, social media activities, and the impact on sleep and grades. Data
from many countries and different education levels were collected. This data was further processed
to handle missing values and convert categorical data into numerical form. The addiction
score was divided into three groups: low, medium, and high. The data was scaled for features
and split into 80% for training and 20% for testing to prepare for model training. The analysis
involved models such as SVM, XGBoost+CNN, CNN+SVM, MLP, and vision transformer.
CNN+SVM and MLP achieved the highest accuracy (0.9220) with training times of 24.23 and
11.58 seconds. Notably, the highest training accuracy was observed with MLP (0.9263), surpassing
other models. Combining multiple layers in these hybrid approaches helped process
the data more effectively and extract key features from the tabular information. Hyperparameter
optimisation improved the performance of ML, DL, and hybrid models. The results show
that ensemble or hybrid models perform better than standalone DL models. The study highlights
how ML and DL can help identify students who may struggle academically and suggest
ways to support them. For future research, larger datasets or real-time data could be used to
improve the models and address some current limitations.
