Recurrent Naive Bayes for Multi-Criteria Recommender Systems: A Novel Approach for Partial Preference Imputation
This study aims to address the issue of partial preference, which causes incomplete criteria-level ratings and limits the accuracy of recommendations in multi-criteria recommender systems (MCRS). The primary focus of this study is to develop an imputation method that treats the imputation process as a sequential prediction task in which each missing criterion is predicted based not only on the known ratings but also on the previously imputed values. These preceding ratings serve as contextual input for the next prediction step, enabling the model to dynamically capture inter-criteria dependencies and thereby improve the overall imputation quality.
A key challenge faced by MCRS is the issue of partial preference, which arises when users fail to provide ratings for all criteria. This results in incomplete criteria-level data, ultimately undermining the accuracy and effectiveness of MCRS in generating relevant recommendations. While most studies have primarily focused on improving the recommendation algorithms themselves, the pre-recommendation phase, which involves preparing high-quality and comprehensive input data, has often been overlooked. Yet, ensuring that multi-criteria data is as complete and accurate as possible is crucial for producing high-quality recommendations. This highlights the importance of a preprocessing step, particularly the imputation process to handle incomplete criteria ratings, applied prior to the recommendation generation process.
Rating-based imputation offers efficiency, while machine learning-based imputation offers accuracy. To leverage those advantages, this study employed rating-based imputation using a simple yet effective machine learning technique, Naive Bayes (NB). NB operates under the assumption of conditional independence among features, which may lead to suboptimal performance in contexts where semantic or perceptual correlations exist among criteria. In response to this limitation, this study introduces a sequential imputation approach, where each prediction of a missing criterion is informed by previously known or imputed ratings that are used as contextual input for the next prediction.
This study introduces a new imputation method called Recurrent Naive Bayes (RNB), designed to estimate missing criteria ratings. Unlike traditional approaches, RNB models the imputation as a sequential prediction process, where each known or already imputed criterion Rc serves as contextual information for predicting the subsequent missing criterion Rc+1. The core of RNB involves a recurrent process that includes prediction using NB, conditional imputation for the missing values, and updating the set of completed criteria. At each step, the model uses more features, starting with the overall rating and adding each newly predicted criterion one by one until all criteria are completed. In the recommendation process, RNB is used beforehand as part of the Multi-Criteria Collaborative Filtering (MCCF) pipeline. The performance of RNB is then evaluated by examining its impact on recommendation accuracy, specifically in terms of predicted criteria ratings, overall ratings, and recommended items, using three real-world datasets: TripAdvisor (TA), Yahoo! Movies (YM), and BeerAdvocate (BA).
This study addresses the issue of partial preference in MCRS by highlighting the underexplored potential of NB for imputing missing criteria ratings. To this end, we propose RNB, a novel method that enhances traditional NB by adopting a sequential imputation strategy. Unlike conventional approaches that treat each criterion independently, RNB captures inter-criteria correlations, thereby improving imputation quality while preserving computational efficiency.
RNB consistently outperformed baseline methods significantly, including no imputation, mean imputation, and NB imputation, in enhancing MCRS performance across all datasets. This is demonstrated by its performance in predicting criteria ratings, overall ratings, and recommended items. The improvements were most notable in datasets with a larger number of criteria, stronger inter-criteria correlations, and larger size. Additionally, RNB enhanced recommendation accuracy for both moderately and highly recommended items.
By implementing MCRS, practitioners are encouraged to incorporate imputation techniques, such as RNB, during the preprocessing stage to improve data completeness and overall recommendation quality. RNB offers a practical balance between accuracy and efficiency, making it suitable for real-world applications where computational resources and data quality vary. It is especially recommended in scenarios involving sparse multi-criteria data with high inter-criteria correlation, such as in tourism, e-commerce, or entertainment domains.
Adopting RNB can be beneficial for researchers as a practical and effective imputation method in MCRS, particularly when dealing with sparse and partially filled datasets. Its ability to capture inter-criteria dependencies through a sequential process makes it a valuable alternative to traditional imputation methods that assume feature independence. RNB’s simplicity, efficiency, and ease of integration into existing MCRS frameworks make it suitable for empirical studies that require scalable and interpretable preprocessing techniques.
Recommender systems play a vital role in supporting decision-making across various domains such as tourism, entertainment, e-commerce, and education. By addressing the issue of partial preferences, RNB improves the completeness of user data, leading to more relevant and personalized recommendations. This impacts enhanced user satisfaction, better overall experiences, increased trust in digital platforms, and more informed consumer decisions. Moreover, the efficiency and simplicity of RNB support broader accessibility by enabling effective recommendations even in data-sparse environments.
The performance of RNB depends on the assumption that inter-criteria dependencies can be captured in a sequential manner. However, the optimal ordering of criteria may differ between datasets, requiring dataset-specific analysis. Future research may explore dynamic or hybrid ordering strategies that simultaneously consider both sparsity and correlation to improve imputation quality. In addition, incorporating more advanced machine learning techniques could help model complex or nonlinear relationships among criteria. Furthermore, enhancing the robustness and adaptability of the imputation process.
Back