IMPACT OF SOCIAL MEDIA ON MENTAL HEALTH USING DEEP LEARNING

Abstract

The study evaluates how deep learning can be used to forecast mental health problems using information on social media taking into account the specificity of speech, behaviors, and other multimodal data inherent to the online media platforms. The complete dataset collected at Kaggle with diverse demographic and social media usage underwent thorough preprocessing, involvement of cleaning, normalization and tokenization along with the state-of-the-art feature extraction tools such as word embeddings and transformers-based representations. The paper presented several deep learning architectures Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, as well as transformer models such as BERT, which have been trained and evaluated on various measures, such as their accuracy, precision, recall, F1-score, and AUC-ROC. Quantitative by their designs, transformer-based and hybrid CNN-Transformer models were outperforming other models consistently at an accuracy rate of over 91, with a precision and recall of over 90 and an F1-score of over 0.90, which is indicative of a stronger capacity to detect contextual and lingo-based cues concerning mental conditions. In comparison, CNN and LSTM models acted more mediocrely, boasting accuracy of 85-87 percent, lower precision and recall figures of 83-87, which underline the importance of modelling context of unstructured social media text. The results also show that the deep learning models will soon provide a powerful 10-15 percent improvement on the predictive accuracy compared to the traditional machine learning baselines, which were Support Vector Machines and Logistic Regression, with an average prediction accuracy of 75-82 percent. Confusion matrices showed that more complex models significantly reduce false positives and false negatives that result in more credible models that are critical in clinical and social health applications. These findings indicate that deep learning-based automated mental health monitoring and intervention systems can be practiced in real-time, particularly multimodal data and longitudinal behavior integration to boost the power of prediction. The thesis offers a fresh view of scalable and decipherable AI models to be used in different contexts of the social media that will become the foundations of the future research of privacy-sensitive algorithms and as an extension of the analyses to the fusion of the audio-visual modality as a continuation of the later improvements of the mental health surveillance.