A Movie Rating Prediction Method Based on Structured Prior Integration and Target-Aware User Interest with Ridge-Based Calibration
- 1Department of Film Press and Contents, Graduate School, Cheongju University, 298 Daeseong-ro, Cheongwon-gu, Cheongju 28503, Chungcheongbuk-do, Republic of Korea
- 2School of Computer Science, Xianyang Normal University, Xianyang 712000, China
Abstract
This study examines whether integrating structured priors with target-aware user histories improves movie rating prediction beyond supervised calibration. CDIAC combines a CatBoost prior, a target-aware Deep Interest Network (DIN) fusion branch, and ridge-based calibration (AC). A chronological MovieLens 32M protocol separates model selection, calibration, and testing and uses temporal out-of-fold priors; calibration coefficients remain fixed during inference. Across five runs on 148,738 warm-start test ratings, the model with a 20-epoch training cap achieves RMSE 0.762500 ± 0.000655 and MAE 0.570105 ± 0.000635. Compared with XGBoost + AC using identical calibration labels, the paired RMSE difference is −0.001500, with a user-cluster bootstrap 95% interval of [−0.002834, −0.000195]. Training and inference require approximately 35.09 and 1.49 times as long, respectively. Component ablations support predictive contributions within the tested pipeline, without establishing DIN-specific superiority. AC reduces aggregate RMSE but increases absolute mean signed error in all four rating groups and pooled RMSE for ratings ≤ 2. These findings indicate a modest accuracy gain at substantially higher training cost. The evidence is limited to the evaluated temporal split and warm-start population: 97.87% of test ratings involve users with at least 50 prior interactions. Feature importance and SHAP characterize only the CatBoost branch.
Related
- Hierarchical Utility Calibration for Structured Multiclass Decisions
- Structured Neural SDEs for Functional Calibration
- Finite-Sample Calibration and Power Benchmarks in Two-Sample Random-Integration Tests
- Multi-Variable Conformal Prediction: Optimizing Prediction Sets using Multiple Calibration Variables
- When Does Trimming Help Conformal Prediction? A Retained-Law Diagnostic under Calibration Contamination
BibTeX
@article{liu2026movie,
title = {{A Movie Rating Prediction Method Based on Structured Prior Integration and Target-Aware User Interest with Ridge-Based Calibration}},
author = {Xiao Liu and Yuyu Liu and Hong Li},
journal = {Mathematics},
year = {2026},
month = nov,
doi = {10.3390/math14213979},
url = {https://www.mdpi.com/2227-7390/14/21/3979},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A Movie Rating Prediction Method Based on Structured Prior Integration and Target-Aware User Interest with Ridge-Based Calibration
AU - Xiao Liu
AU - Yuyu Liu
AU - Hong Li
PY - 2026
DA - 2026/11/01
JO - Mathematics
DO - 10.3390/math14213979
UR - https://www.mdpi.com/2227-7390/14/21/3979
AB - This study examines whether integrating structured priors with target-aware user histories improves movie rating prediction beyond supervised calibration. CDIAC combines a CatBoost prior, a target-aware Deep Interest Network (DIN) fusion branch, and ridge-based calibration (AC). A chronological MovieLens 32M protocol separates model selection, calibration, and testing and uses temporal out-of-fold priors; calibration coefficients remain fixed during inference. Across five runs on 148,738 warm-start test ratings, the model with a 20-epoch training cap achieves RMSE 0.762500 ± 0.000655 and MAE 0.570105 ± 0.000635. Compared with XGBoost + AC using identical calibration labels, the paired RMSE difference is −0.001500, with a user-cluster bootstrap 95% interval of [−0.002834, −0.000195]. Training and inference require approximately 35.09 and 1.49 times as long, respectively. Component ablations support predictive contributions within the tested pipeline, without establishing DIN-specific superiority. AC reduces aggregate RMSE but increases absolute mean signed error in all four rating groups and pooled RMSE for ratings ≤ 2. These findings indicate a modest accuracy gain at substantially higher training cost. The evidence is limited to the evaluated temporal split and warm-start population: 97.87% of test ratings involve users with at least 50 prior interactions. Feature importance and SHAP characterize only the CatBoost branch.
SN - 2227-7390
ER -