Accurate and interpretable movie rating prediction remains challenging. Many existing models make only partial use of narrative text, treat heterogeneous ratings with simple noise assumptions, and offer limited insight into their internal decisions. To address these issues, this paper proposes the Narrative-Aligned Multimodal Rating Network (NAMRN), a model that jointly exploits plot descriptions and structured movie attributes while explicitly modeling sample-wise uncertainty. NAMRN consists of three main components. A narrative-aligned contrastive pretraining module learns plot-level representations that are closely tied to rating signals. An uncertainty-aware heteroscedastic regression module predicts both the mean rating and its variance, so that samples with different confidence levels are treated differently in the loss. A sparse-gated multimodal fusion module adaptively selects informative features from textual and structured channels, which reduces redundancy and highlights the most relevant attributes. All components are compatible with gradient-based interpretability methods, which allows detailed inspection of token-level and feature-level contributions. Experiments on three public movie datasets demonstrate the effectiveness of NAMRN. On The Movies Dataset, NAMRN achieves an MAE of 0. 124, an RMSE of 0. 170, and an R² of 0. 81, outperforming support vector regression, gradient boosting, LSTM-based models, and Transformer-based baselines. Ablation studies confirm that removing any of the three core modules leads to consistent performance degradation. Cross-dataset evaluation on IMDb and MovieLens further shows that the model maintains stable error levels under different rating scales and data distributions. Interpretability analysis based on gradient heatmaps reveals that NAMRN focuses on meaningful narrative cues and structured attributes that are consistent with human judgement. Overall, NAMRN provides a unified solution that improves accuracy, robustness, and transparency in movie rating prediction. The proposed design also offers a flexible foundation for future extensions to multi-criteria evaluation, additional modalities, and fairness-aware recommendation.
Peng et al. (Tue,) studied this question.