Enhancing the Reliability of Affect Recognition in Social Platforms with Conformal Prediction

Reliable affect recognition is essential in computational social systems, particularly when analyzing user-generated content on social platforms, where accurate emotion interpretation supports effective content moderation, social analysis, user modeling, and crisis detection. Although large language models (LLMs) achieve strong affect-recognition performance after task-specific fine-tuning, they often remain poorly calibrated, frequently assigning high confidence to incorrect predictions and lacking reliable uncertainty estimates. Such miscalibration undermines the trustworthiness, interpretability, and practical utility of socially intelligent systems. This study applies conformal prediction (CP), including split and adaptive variants, as a model-agnostic post hoc framework that provides finite-sample coverage guarantees and statistically valid uncertainty quantification. We present a unified CP framework that extends existing approaches by developing global, Mondrian, and hybrid conformal predictors for ordinal and multilabel affect recognition, as well as adaptive CP via conformalized quantile regression for regression tasks. Extensive experiments on datasets derived from user-generated social content demonstrate that CP achieves reliable coverage, independent of task type, evaluation metric, or underlying LLM architecture.