Asian Journal of Research in Computer Science

Aims/Objectives: Traditional anime recommendation systems frequently rely on viewing history, user ratings, and genre-based filtering, all of which often fail to capture the complex emotional immersion inherent in the medium. Given that emotional resonance is a primary driver of audience engagement, incorporating affective data can significantly enhance recommendation accuracy.

Study Design: Quantitative experimental research with hybrid machine learning framework.

Methodology: This study proposes an emotion-aware hybrid anime recommendation system that extracts emotional patterns from dialogue using Natural Language Processing (NLP). The methodology involves preprocessing over 1.5 million anime subtitle lines sourced from a large-scale English-translated anime subtitle corpus, covering thousands of episodes across diverse genres, to eliminate noise and segment dialogue for fine-grained analysis. A transformer-based emotion detection model (j-hartmann/emotion-english-distilroberta-base) is applied to each dialogue line to generate probability distributions across seven emotion categories. These predictions are aggregated to produce episode-level emotion vectors, which are further augmented with three engineered "anchor features" — Total Intensity, Emotion Breadth, and a Positivity Index — expanding the input representation from a 6-dimensional (6D) baseline to a 9-dimensional (9D) feature space.

Results: Six hybrid regression models were benchmarked against both the 6D baseline and 9D feature configurations under Gaussian noise injection (0–10%), simulating real-world variability in user emotional self-reporting. The 9D anchor feature integration produced a 33% to 65% improvement in Noise Resilience Score (NRS) across all architectures compared to the 6D baseline, with attention-based models (TabPFN, TabNet) recovering from near-complete failure (negative R² at 10% noise in the 6D system) to robust performance (R² = 0.7227–0.7792). XGBoost emerged as the optimal backbone with the highest NRS of 0.8980 and R² = 0.7999 under maximum noise, a finding independently validated by the AutoML framework across 141 search iterations. The hybrid scoring pipeline, combining the 9D XGBoost regression model with a Sentence-Transformer semantic layer (all-MiniLM-L6-v2), demonstrated approximately 80% inter-model agreement across architectures for identical user inputs.

Conclusion: This approach yields suggestions that are more context-aware, psychologically grounded, and aligned with the emotional expectations of viewers, establishing a technically robust foundation for future multimodal affective recommendation systems.