Emotion-conditioned podcast recommendation using context-aware collaborative filtering

Abstract

The rapid growth of podcast consumption on video-based platforms introduces new challenges for recommender systems, as user media choices are influenced not only by historical preferences but also by dynamic psychological states. Conventional approaches primarily model similarity and often ignore emotional context, particularly in long-form media. This study investigates emotion-conditioned recommendation behavior by treating user emotion as a contextual variable within a hybrid recommendation framework. This framework models emotion as a conditioning mechanism that constrains candidate selection prior to collaborative ranking. A dataset of 4,468 YouTube podcast transcripts was collected and preprocessed. Emotional labels were generated using the NRC emotion lexicon and validated through manual verification. User emotional state was detected using a support vector machine (SVM), while implicit preferences were estimated from engagement indicators using a random forest (RF) model. The recommendation stage integrates contextual pre-filtering based on Plutchik’s emotional relationships with collaborative filtering using k-nearest neighbors (KNN) with cosine similarity. Evaluation using stratified 5-fold cross-validation, baseline comparison, and Wilcoxon signed-rank testing shows that emotional context alters recommendation ranking behavior and improves ranking quality and retrieval coverage within the candidate space. These findings indicate that emotion acts as a conditioning mechanism in long-form media consumption, influencing recommendation outcomes beyond predictive accuracy.