DOI: 10.1177/09731342261471649 ISSN: 0973-1342

Pilot Study of Machine Learning-based Emotion Analysis (Ages 9–14): Interpretable Approaches to Schadenfreude and Sympathy Classification

Sonam Sultana, Bijoy Krishna Panda, Muktipada Sinha

Background:

Children’s emotional responses to others’ misfortunes, particularly schadenfreude and sympathy, play a pivotal role in social development, influencing peer relationships and moral reasoning. Traditional assessments often rely on self-report measures, which are vulnerable to cognitive biases and social desirability. To address these limitations, this pilot study introduces a multimodal elicitation framework integrating scenario-based interviews, illustrated stories, video stimuli, and observational validation to capture both explicit and implicit emotional cues. As a pilot study, this research primarily tested the feasibility and interpretive validity of a multimodal framework rather than drawing definitive generalizations.

Method:

Four machine learning (ML) classifiers—Decision Tree (DT), Random Forest (RF), Logistic Regression (LR), and Gaussian Naïve Bayes—were applied to categorize children’s moral-emotional responses. The supervised models predicted internally derived latent emotional profiles from k -means analysis, rather than external or clinically validated ground truth labels. k -Means clustering revealed distinct emotional profiles, validated through Kruskal-Wallis testing. SHAP and Gini-based feature importance analyses identified key predictors driving classification, particularly scenario-based inputs from video stimuli.

Results:

LR achieved the highest overall performance (Accuracy = 0.97, Precision = 0.97, Recall = 0.97, F 1 = 0.97), followed by GNB and RF, with all three outperforming DT. Scenario-based features from morally charged contexts (Videos 3 and 4) emerged as the strongest predictors. Observational validation confirmed alignment between behavioral cues and predicted emotional profiles, enhancing the study’s ecological validity.

Conclusion:

This research demonstrates the value of interpretable ML in classifying morally ambivalent emotions in children. The framework was designed with Generation Z learners in mind, as they are increasingly exposed to emotionally complex content through social media and digital platforms. The proposed framework has the potential to evolve into a cautiously scalable framework, subject to validation across larger and more diverse educational settings. It could enable educators to identify students’ emotional tendencies early, support empathy development, and tailor social-emotional learning interventions in real time.

More from our Archive