An Explainable Counterfactual and Agent-Assisted Machine Learning Framework for Emotion-Aware Digital Mental Well-Being Decision Support

An explainable, prescriptive machine learning framework for digital mental well-being. It predicts a person's dominant emotional state from how they use social media, explains that prediction for the individual and across the whole dataset, finds the smallest behaviour changes that would move a negative state toward a positive or stable one, and turns the result into real-time guidance through a rule-based agent.

Seven behavioural features (age, platform, daily usage time, and posts, likes, comments and messages per day) are used to predict a user's dominant emotion across seven classes. The platform is encoded as a number, and an ANOVA test confirms every feature is significant (p < 0.05), with likes and comments received showing the largest effect.

The classes were highly imbalanced, so Random Oversampling brought every class to 257 samples before a 60/40 train-test split. An optimised Random Forest is the predictive core. LIME explains individual predictions and SHAP gives global feature importance.

A DiCE counterfactual module then searches for the closest feature vector the model classifies as Happiness or Neutral, with every feature kept inside its range and age allowed only to stay the same or rise (up to 60). Finally, an Agentic AI module applies 15 decision rules on the predicted emotion, age, platform, usage and engagement to produce personalised recommendations through a simple input form.

The optimised Random Forest reached 0.98 overall accuracy on the 40% held-out test split, with per-class F1-scores between 0.96 and 1.00. ROC AUC was 1.000 for four of the seven classes and no lower than 0.995 for the rest, and average precision was close to 1.000 for every class. The calibration curves stayed close to the diagonal, with small deviations for Anxiety and Sadness, so the confidence scores are a reasonable input for the recommendation layer.

SHAP ranked comments received per day as the most influential feature overall, followed by daily usage time and likes received; platform and posts per day mattered least. The DiCE case studies gave concrete recourse: for example, a 33-year-old Instagram user predicted as Aggression reaches Neutral with fewer posts (6 to 1), fewer messages (29 to 19) and slightly less daily use (140 to 135 minutes). The agent module produced matching recommendations for all seven emotional states.

Structured social media behaviour can predict emotional well-being states accurately, and pairing the prediction with LIME and SHAP explanations, DiCE counterfactual recourse and a rule-based agent turns a black-box classifier into a transparent, prescriptive decision-support tool. Its outputs are decision-support insights, not clinical diagnoses. Future work targets larger, longitudinal and multi-modal datasets, causal explanations, and recommendations that adapt to user feedback.