Supervised Learning for Emotional Prediction and Feature Importance Analysis Using SHAP on Social Media User Data.

Hikmawati, Erna; Alamsyah, Nur

This study aimed to predict emotional states using supervised learning models and analyze the importance of features in social media user data. We implemented the Random Forest algorithm to predict happiness, neutrality, and sadness based on various social media activity metrics, including daily usage time, posts per day, and interactions such as likes and comments received. Data preprocessing involved handling missing values, coding categorical features using One-Hot Encoding, and scaling numerical features with StandardScaler. We assessed the model's performance utilizing Mean Squared Error (MSE) and R-squared (R²) measures. The results showed that the model had a high prediction accuracy, with R2 values of 0.897 for happiness, 0.863 for neutrality, and 0.851 for sadness. SHapley Additive exPlanations (SHAP) were used to perform a thorough feature importance analysis, which revealed that daily usage time and user interaction significantly influenced emotional states. These findings underscore the efficacy of combining supervised learning with SHAP for interpretable and accurate emotional predictions, providing valuable insights for the development of tools and strategies to monitor and enhance emotional health in the digital era.

RANDOM forest algorithms; DIGITAL technology; SUPERVISED learning; EMOTIONAL state; MENTAL health

Ingénierie des Systèmes d'Information, 2024, Vol 29, Issue 6, p2345

1633-1311

Academic Journal

10.18280/isi.290622