TY  - EJOU
AU  - Zheng, Yong 

TI  - A Comprehensive Review of Rating Imputation in Recommender Systems: From Data Completion to Inference-Oriented Missing-Data Estimation
T2  - Computers, Materials \& Continua

PY  - 
VL  - 
IS  - 
SN  - 1546-2226

AB  - Recommender systems can alleviate information overload by producing item recommendations tailored to user preferences. The performance usually relies on rich user-item interaction data; however, missing entries introduce sparsity that substantially degrades performance. Early work primarily treated rating imputation as a preprocessing mechanism for mitigating sparsity and alleviating cold-start issues through explicit matrix completion. More recently, missing-data estimation has evolved beyond static preprocessing toward broader inference-oriented paradigms, including pseudo-label estimation, counterfactual inference, and debiasing mechanisms integrated directly into the learning objective. In this paper, we present a structured review of rating imputation and inference-oriented missing-data estimation methods in recommender systems, organized through a conceptual perspective rather than a single formal framework. We organize existing approaches into four stages: statistical approaches, machine learning-based techniques, methods leveraging auxiliary information, and recent advances grounded in causal inference and neural approaches (e.g., deep learning and large language models). Beyond sparsity reduction, we emphasize the role of missing-data estimation in emerging contexts, including correcting selection bias under Missing Not At Random data, inferring partial preferences in multi-criteria recommendation systems, and integrating heterogeneous rating sources. To address an underexplored aspect of the literature, we further analyze potential limitations of imputation and inference-oriented estimation, such as covariance distortion, overconfidence, poisoning effects, and robustness degradation under inaccurate pseudo-labels. By introducing a structured taxonomy, this survey provides a systematic and conceptually organized perspective on missing-data handling in recommender systems and highlights the importance of rigorous bias analysis and robustness evaluation frameworks for developing reliable and unbiased recommendation systems.
KW  - Recommender system; missing data; rating imputation; causal inference; bias correction

DO  - 10.32604/cmc.2026.084278