In recent decades, the global seafood trade has expanded dramatically, with international seafood imports reaching unprecedented volumes to meet growing consumer demand worldwide [1]. This expansion has been accompanied by increasing concerns regarding food safety, as seafood products are particularly susceptible to various contaminants, including heavy metals, pathogens, and unauthorized additives [2]. Regulatory authorities around the world have implemented rigorous inspection systems at entry points to ensure that imported seafood meets established safety standards. However, the effectiveness of these inspection systems faces significant challenges due to the vast volume of imports and the relatively low incidence of non-conformity, approximately 0.2%, as observed in our dataset.

Food safety is deeply affected by the diversity of foods and their raw materials. The increasing trade openness of the global economy has resulted in increased food imports, emphasizing the significance of food risk management in protecting consumer health. Predictions and early warning are crucial to ensure food safety; in particular, food inspection prior to entry into the consumer market is a significant step in ensuring good food quality.

In this context, the term non-conformity is used to denote inspection failures resulting from violations of established safety and quality criteria. These include microbiological hazards, chemical residues, and labeling defects, which are evaluated according to the regulatory standards enforced by the Korean Ministry of Food and Drug Safety [3].

The identification of key factors influencing non-conformity in imported seafood represents a critical area for research, as it could potentially enhance the efficiency and effectiveness of inspection protocols. Previous studies have examined various aspects of seafood safety, including the prevalence of specific contaminants [4, 5], geographical variations in compliance rates [6], and seasonal fluctuations in detection rates [7]. However, there has been little research on the use of proactive inspections for high-risk food predictions as part of the border control for imported foods, with the exceptions of the United States (US) and the European Union (EU). Furthermore, there remains a significant gap in understanding how multiple factors—such as seafood type, distribution method, import timing, country of origin, and characteristics of importers and exporters—collectively influence the likelihood of non-conformity. Furthermore, governmental implementations such as the risk prediction-based imported food inspection system developed by the Korean Ministry of Food and Drug Safety [8] have established procedural precedents through logistic modeling approaches incorporating multidimensional risk factors, though such systems exhibit methodological constraints when addressing extreme classification imbalance and provide limited explainability mechanisms for specialized product categories.

The inherent challenge in developing predictive models for seafood inspection outcomes lies in the severe class imbalance within the available data. With non-conformity rates of approximately 0.2% (796 non-conforming cases from 389,389 total observations in our study), traditional classification algorithms often fail to accurately identify the minority class (non-conforming samples), instead favoring the overwhelming majority class (conforming samples) [9]. This imbalance problem is particularly problematic in regulatory contexts where the cost of missing a non-conforminity product (false negative) significantly outweighs the cost of additional inspection for a compliant product (false positive).

Despite several studies focused on using machine learning for predicting the results of inspections, there is a lack of information regarding the model decision and explainability, such as that found in the areas of water quality [10, 11] and healthcare [12, 13]. To address this gap, our study employs SHapley Additive exPlanations (SHAP) [14] to provide principled and interpretable insights into the model’s decision-making process.

This study addresses these challenges by developing and evaluating multiple classification models designed to predict non-conformity in imported seafood products while effectively handling the class imbalance problem. Specifically, we implement and compare four distinct classification algorithms—Decision Tree (DT), Random Forest (RF), Logistic Regression (LR), and Naive Bayes (NB)—and employ both hard and soft voting ensemble mechanisms to determine the most effective approach. Our methodology incorporates specialized data preprocessing techniques to mitigate the effects of class imbalance, including strategic sampling methods and feature engineering tailored to the unique characteristics of seafood import data.

The primary objectives of this research are: (1) to identify the key factors that significantly influence non-conformity in imported seafood products; (2) to develop an effective classification model that accurately identifies potentially non-conformant products despite the severe class imbalance; and (3) to provide actionable insights for regulatory authorities to optimize inspection protocols and resource allocation. By achieving these objectives, this study aims to contribute to the enhancement of food safety systems while simultaneously reducing unnecessary inspection burdens on compliant imports.

The results demonstrated that a soft voting ensemble approach achieved superior performance in identifying non-conformity cases, with a recall of 75.57% and Area Under the Curve (AUC) of 87.49%, outperforming the hard voting method. SHAP analysis confirmed that exporting country ratio, major product category, manufacturer ratio, importer ratio, and seasonal variations significantly influenced model decisions providing comprehensive analysis by considering multiple risk dimensions and seasonality. This research not only enhances seafood inspection efficiency but also offers a methodological framework that can be readily applied to other product categories within the food sector facing similar class imbalance challenges.

The remainder of this paper is organized as follows. Section 2 provides a review of related literature on data sampling methods, ensemble learning, and explainable artificial intelligence. Section 3 describes the materials and methods, including data sources, preprocessing, model construction, and evaluation. Section 4 presents the experimental results and interpretability analysis. Section 5 discusses the implications of the findings. Finally, Section 6 concludes the study and outlines directions for future research.

This section presents a comprehensive overview of the theoretical foundations and prior research relevant to the prediction of food inspection outcomes. It introduces essential concepts including data sampling strategies for addressing class imbalance, ensemble learning techniques with an emphasis on voting mechanisms, and explainable artificial intelligence (XAI).

The objective is to establish a solid conceptual and methodological framework for the approaches employed in this study. Furthermore, this section identifies existing research gaps, particularly the limited application of interpretable machine learning models in the domain of food safety inspections, thereby highlighting the novelty and necessity of this study.

2.2 Voting Ensemble Model

A voting ensemble is a machine learning ensemble model that combines predictions from multiple models. This technique can be used to improve the model performance, ideally outperforming any single model in the ensemble. A voting ensemble combines the predictions from multiple models. This method is suitable for classification; during classification, the predictions for each label are added, and the label with the most votes is predicted [19].

Two approaches are available to predict the majority votes for classification, namely hard voting and soft voting. As shown in Fig. 1, hard voting entails adding up all the predictions for each class label and predicting the class label with the most votes. Meanwhile, soft voting averages the predicted probabilities for each class label and predicts the class label with the greatest probability [20].

Fig. 1.

Voting ensemble techniques explanation. Comparison of hard and soft voting ensemble techniques where hard voting uses majority rule while soft voting averages prediction probabilities across multiple classifiers.

This section presents the methodological framework employed to develop and assess the machine learning models for predicting seafood product import inspection outcomes as illustrated in Fig. 2. The study process is described in detail, including data acquisition, preprocessing procedures, feature engineering, and the construction of classification models. Particular attention is given to handling class imbalance using resampling techniques, selecting statistically significant features, and implementing ensemble learning approaches through hard and soft voting mechanisms. Furthermore, model evaluation metrics and explainable AI techniques are introduced to ensure both the reliability and interpretability of the predictive results.

Fig. 2.

Workflow of the four-phase methodology. The study process includes four phases. Machine learning models used in the study are Decision Tree (DT), Random Forest (RF), Logistic Regression (LR), and Naive Bayes (NB). AUC, Area Under the Curve; XAI, explainable artificial intelligence.

3.3 Model Construction

3.3.1 Feature Selection

The preprocessed data were subjected to a two-stage variable selection process, in which attributes were designated for inclusion in the model construction. In the first stage, a single-factor analysis was performed to identify factors that had statistically significant relationships with conformity or non-conformity during inspections. Different statistical tests were adopted depending on the variable type. The continuous variables, such as the total net weight, were analyzed using the ANOVA test. In contrast, the remainder of the variables, namely the categorical variables, were analyzed using the chi-squared test [22].

3.3.2 Spliting of the Data into Training and Testing Datasets

To acquire the optimal models, perform model validation, and evaluate the model performance, the dataset was split into two groups—80% for training and 20% for testing. After feature selection was constructed, dataset used for modeling were divided into the test and training datasets and were subsequently oversampled to balance the training data as shown in Fig. 3.

Fig. 3.

Flowchart for the prediction model. Data preparation flow for balancing the highly imbalanced seafood inspection dataset through resampling before model training and evaluation.

The main purpose of resampling was to enhance the discriminatory ability of the model rather than to learn erroneous samples. Moreover, the test dataset deviated from the original data if sampling was performed before splitting the data. Consequently, the model learned noise from the data, resulting in inaccurate predictions. The dataset was oversampled using SVM-SMOTE-SVM, and the resampled training dataset was employed during model training to establish the most suitable model for the testing dataset and XAI. The resampled data for training consisted of 497,425 data points for the conform class and 248,603 for the non-conform class.

3.3.3 Modelling

Four types of models, namely, Decision Tree (DT) [23], Random Forest (RF) [24], Logistic Regression [25], and Naive Bayes (NB) [26], were used to create ensemble models for model training. Because non-conformity is considered a minority class, this study used cost-sensitive learning that considers the costs of different misclassifications [27]. Using the balanced method, the class weights were inversely proportional to the class frequencies in the training dataset. Using class weights, the model learned to minimize the total cost, not just the number of misclassifications. This can be beneficial in situations where misclassification costs are unevenly distributed. As mentioned above, two types of ensemble models were used in this study, namely soft and hard voting models. The performances of the soft and hard voting models were compared using both the resampled validation and the test datasets. XAI was then used to check the feature importance of each model (i.e., DT, RF, LR, and NB).

The seafood inspection classification model presented in Fig. 4 constitutes a taxonomic framework designed to predict binary inspection outcomes through the integration of 27 predictor variables strategically organized into seven distinct categories with the detailed predictors and response variable. These categories—comprising Entity Identification Variables (n = 4), Product Characterization Variables (n = 3), Quantitative Assessment Variables (n = 2), Logistical-Procedural Variables (n = 2), Temporal-Chronological Variables (n = 6), Geographical Variables (n = 1), and Ratio Metrics (n = 9)—establish a comprehensive analytical foundation for the voting classifier mechanism. The model employs SHAP analysis to derive feature importance rankings through both global assessments and local explanations, thereby facilitating interpretability of the classification outputs. A representative sample from our dataset is presented in Supplementary Table 1, illustrating the 27 predictor variables used in the classification model.

Fig. 4.

Seafood inspection classification model. Illustrating the taxonomic organization of predictor variables and their relationship to the binary classification outcome. SHAP, SHapley Additive exPlanations.

3.3.4 Model Evaluation

The model performance were measured and validated using a confusion matrix and model predictive performance indicators to select the optimal model and evaluate its performance. A confusion matrix was structured using the entries listed in Table 2, and the necessary predictive performance indicators were calculated using the numbers of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN).

Table 2. Definitions of entry types in the confusion matrix.

Entry type	Definition
True Positive (TP)	Predicted inspection result for the product by model classification: non-conformity; actual inspection result: non-conformity.
False Positive (FP)	Predicted inspection result for the product batch by model classification: non-conformity; actual inspection result: conformity.
True Negative (TN)	Predicted inspection result for the product batch by model classification: conformity; actual inspection result: conformity.
False Negative (FN)	Predicted inspection result for the product by model classification: conformity; actual inspection result: non-conformity.

Predictive performance indicators included the area under the curve (AUC), the positive predictive value (PPV) (also known as precision), the F1 score, the recall, and the accuracy (ACR), which are defined in detail below.

• The accuracy rate (ACR) evaluates the model’s overall capacity to differentiate between conformity and non-conformity samples or the ability to accurately classify samples as conformity. However, owing to the lower proportion of non-conformities in our data, there was an imbalance in the samples considered herein. Because of its higher capacity for discriminating conformities, ACR may show bias in predicting conformities. To overcome this problem, the recall and PPV indicators have received greater attention during the evaluation of model performance. The ACR can be calculated using (2):

(2)$$ACR=\frac{TP+TN}{TP+TN+FP+FN}$$

• The recall or sensitivity is the proportion of samples correctly labeled as non-conformity out of all non-conformity samples, as shown in (3):

(3)$$Recall=\frac{TP}{FN+TP}$$

• The positive predictive value (PPV), also known as precision, is the proportion of samples that the model classifies as non-conformity out of all samples, and is otherwise referred to as the non-conformity rate. The PPV can be calculated using (4):

(4)$$PPV=\frac{TP}{TP+FP}$$

• The F1 score, defined as the harmonic mean of the recall and PPV indicators, becomes crucial when dealing with imbalanced data. Higher TP values correlate with higher F1 scores, and the F1 score can be calculated using (5):

(5)$$F1=\frac{2\cdot PPV\cdot Recall}{PPV+Recall}$$

• The model’s classification accuracy can be measured from the area under the receiver operating characteristic (ROC) curve (AUC), wherein a larger AUC denotes a higher accuracy. More specifically, AUC = 1 represents a great classifier, 0.5 < AUC < 1 represents a model that outperforms random guessing, AUC = 0.5 represents a model that is similar to random guessing but lacks classification capacity, and AUC <0.5 represents a classifier that performs worse than random guessing.

According to the explanation above, recall and AUC scores play an important role in the model evaluation. The higher scores show that the higher chance model can correctly identify the non-conformity class.

This study addressed the significant challenge of predicting non-conformity in imported seafood products—a critical task in food safety management characterized by severe class imbalance. The methodological framework integrated ensemble learning techniques with explainable AI approaches to develop robust prediction models while maintaining interpretability. The empirical findings warrant comprehensive discussion across several dimensions.

Four different models were used, namely the NB, DT, RF and LR models, which were stacked together using class-weight cost-sensitive learning to create both hard-voting and soft-voting ensemble models. These models were applied to forecast nonconformance after training, and their performances were evaluated by using the test dataset. Table 3 summarizes the performance metrics of the models. Among the two ensemble methods, soft voting outperformed hard voting in terms of the recall score, receiving 75.57% of votes compared to the hard voting method’s 44.32%. Additionally, the soft voting method achieved a higher AUC score of 87.49%, whereas the hard voting method obtained a score of 72.07%. However, hard voting achieved better results in terms of the ACR (99.69%, c.f., 99.35% for soft voting). Similarly, hard voting outperformed soft voting in the PPV and F1 scores, with scores of 35.62 and 39.49%, respectively; soft voting obtained scores of 22.32 and 34.46%, respectively. These findings suggest that the soft voting method performed better than the hard voting method in terms of the recall and AUC scores, thereby indicating that the soft voting ensemble is more effective in correctly identifying non-conformity data and predicting inspection outcomes. Thus, the soft voting approach may be more suitable in cases where the recall and AUC scores are more critical than the PPV and ACR scores. This finding aligns with established theoretical frameworks regarding ensemble techniques for imbalanced classification problems while extending empirical validation specifically to the domain of food safety inspection by Douzas et al., and Wu & Weng [9, 29, 30].

To gain insight into the decision-making process of each model in the ensemble, SHAP values were calculated for all models. Additionally, local-level feature contributions to the predictions of four representative samples were analyzed to obtain deeper insights. Global feature importance analysis revealed the predominant influence of historical performance indicators, particularly the exporting country ratio, major product category, overseas manufacturer ratio, and importer ratio. These findings validate existing risk-based approaches to food safety management that prioritize historical compliance patterns as predictors of future regulatory adherence, as demonstrated by Djekic & Jankovic [31] in their analysis of food safety notifications in the European Union. More importantly, these findings resonate with existing risk-based inspection frameworks, such as the system developed by the Korean Ministry of Food and Drug Safety, outlined in patent KR20140077006A [8].

The exporting country ratio emerged as one of the most influential features across all models. This metric captures historical non-conformity rates by country, reflecting differences in regulatory standards and supply chain controls. This finding aligns with Fanani and Yuadi [32], who identified Russia, Mauritania, Papua New Guinea, and Solomon Islands as high-risk exporters to the U.S. Our SHAP analysis confirms that country-specific factors strongly influence non-conformity predictions, indicating that regional disparities in fishing practices and compliance create identifiable risk patterns that can inform targeted inspection protocols.

The overseas manufacturer ratio significantly influences food safety outcomes, highlighting the importance of producer-specific factors in risk assessment. This reflects variations in quality control systems, manufacturing practices, and compliance histories at the facility level. Our local SHAP analysis Fig. 6B of Sample 2 revealed that even products from low-risk countries may present elevated risk when produced by manufacturers with higher non-conformity ratios. This finding supports inspection approaches focused on manufacturer-specific histories rather than relying solely on country-level assessments, consistent with Riviere & Buckley’s [33] research on quality control systems and Suanin [34] findings on the importance of compliance histories in seafood product safety.

The importer ratio demonstrated substantial predictive power, reflecting the critical role of importer compliance history in forecasting inspection outcomes. As shown in Fig. 6A,C, varying importer ratios significantly impacted non-conformity predictions, with higher historical non-conformity rates strongly correlating with increased risk assessment. This pattern indicates that importer-specific factors, such as supplier selection practices, quality management systems, and regulatory compliance commitment, create persistent risk patterns that can be effectively captured through historical performance metrics. The model’s ability to identify these patterns enables regulatory authorities to implement targeted verification protocols for importers with problematic compliance histories while potentially reducing inspection burdens on consistently high-performing entities. This risk-based approach to importer assessment aligns with Bouzembrak & Marvin’s [35] framework for prioritizing inspection resources based on historical compliance data and supply chain characteristics.

The importance of temporal features—such as year, month, and week—in predicting seafood inspection outcomes highlights the presence of seasonal patterns in food safety risks. Seasonal variations, particularly fluctuations in water temperature, can significantly influence microbial activity in seafood. Warmer temperatures during summer months are associated with increased prevalence of pathogens like Vibrio species, a trend supported by Zhang et al. [36]. This seasonal impact is further illustrated in Fig. 6D, where a notable increase in non-conformity predictions is observed around the 28th week, corresponding to the summer season. Besides, seasonal harvesting patterns create periodic surges in processing volumes that can strain quality control systems. Additionally, rainfall variations affect coastal water quality through agricultural runoff, potentially introducing contaminants into harvesting areas during specific seasons [37]. In the local explanation Fig. 6C also identifies June, which is a rainy month, as the influence factor on the non-conformity result.

The major product category emerged as a significant predictor across all models, highlighting the variability in risk profiles among different seafood types. This finding aligns with the established understanding that certain seafood categories inherently pose higher contamination risks due to their biological characteristics and supply chain complexities. For example, filter-feeding bivalves like short-neck clams (Sample 2 in Fig. 6B) are particularly susceptible to environmental contaminants, while products requiring extensive processing may experience quality control challenges. The pronounced influence of product categories on model predictions was consistently observed in the SHAP analysis, with varied effects across different seafood types. This suggests that inspection resources should be strategically allocated based on product-specific risk profiles rather than applying uniform protocols across all seafood imports, an approach supported by FAO [38] research on compliance patterns in processed food exports.

This study addressed the significant challenge of developing a generalizable machine learning model for predicting non-conformity in seafood product importation inspections. By implementing an ensemble approach that combines Decision Trees, Random Forests, Logistic Regression, and Naive Bayes models, the study handled the severe class imbalance (0.2% non-conformity rate) inherent in food safety inspection data.

The soft voting ensemble technique demonstrated better performance in identifying non-conforming seafood products, achieving a recall of 75.57% and an AUC of 87.49%, significantly outperforming the hard voting method (44.32% recall, 72.07% AUC). Through SHAP analysis, the study identified key factors influencing inspection outcomes, including exporting country ratio, major product category, overseas manufacturer ratio, importer ratio, and seasonal variations.

The study findings suggest that historical compliance patterns serve as strong predictors of future regulatory adherence, with particular importance placed on country-specific factors and producer-specific compliance histories. The identification of temporal patterns indicates seasonal variations in food safety risks, likely influenced by environmental factors such as water temperature and rainfall patterns.

Importantly, while this research focused specifically on seafood products, the methodological framework developed here can be readily applied to other product categories within the food sector and potentially beyond. The ensemble learning approach combined with explainable AI techniques provides a good template that can be adapted to predict inspection outcomes for diverse imported goods such as meat products, processed foods, and even non-food consumer goods subject to regulatory oversight. The core principles of leveraging historical compliance data, entity-specific risk factors, and temporal patterns remain applicable across various inspection domains facing similar class imbalance challenges.

The methodology developed in this study offers a practical framework for regulatory authorities to implement risk-based inspection protocols, enhancing both efficiency and effectiveness of limited inspection resources. By targeting high-risk imports based on multiple risk dimensions, authorities can simultaneously improve food safety and reduce unnecessary inspection burdens on consistently compliant entities.

Nevertheless, this study has the following limitations. Firstly, there are limitations in refining potential errors or unstructured items in the data, and some important variables (e.g., details of actual inspection criteria) were not included. Secondly, the ensemble model configuration used only four traditional classifiers, and no comparison was made with recently developed deep learning-based models. Thirdly, while SHAP analysis provided valuable insights into feature importance and local explanations, our implementation did not fully explore advanced SHAP capabilities such as interaction effects and dependence plots that could have offered deeper interpretability of the model’s decision-making process. Lastly, there is a challenge in handling new entities, such as first-time importers, manufacturers, or products without historical data in the predictive framework.

Future research may consider the following directions: (1) linkage of in-depth feature engineering and external data (e.g., climate, supply chain risks, etc.), (2) multi-layer ensemble or hybrid model configuration and performance optimization based on AutoML, (3) empirical evaluation such as policy field application experiments and cost-effectiveness analysis. In particular, follow-up research is needed to maximize policy application performance through integrated scenario design with field inspection systems and linkage with real-time warning systems. (4) Development of transfer learning methods that apply knowledge from similar established entities to new ones. (5) Further exploration of advanced XAI techniques, including more sophisticated applications of SHAP analysis such as interaction values, dependence plots, and distribution analysis across product categories, which could provide deeper insights into model decision-making processes and enhance the interpretability of predictions for regulatory stakeholders.

The data presented in this study are available on request from the corresponding author. The data are not publicly available due to concerns related to the protection of participant privacy and confidentiality.

SK contributed substantially to the conceptualization and design of the study, performed formal analysis, developed the software, prepared the original draft, and created visualizations. KL contributed to the conceptualization of the study and managed the project. WC supervised the research and contributed to critical revision of the manuscript. All authors contributed to manuscript editing, read and approved the final manuscript, and agree to be accountable for all aspects of the work.

Not applicable.

This research was supported by a grant (21163MFDS517-1) from Ministry of Food and Drug Safety.

Wan-Sup Cho is the owner of BigDataLabs Co., Ltd., and Saksonita Khoeurn was employed by BigDataLabs Co., Ltd. The company develops and markets products related to the topic of this manuscript. However, the company had no role in the design, analysis, or interpretation of the study. The authors declare that there are no conflicts of interest.

Supplementary material associated with this article can be found, in the online version, at https://doi.org/10.31083/JFSFQ39242.