From Data to Strategy: Explainable AI Insights for Enhancing Airline Passenger Satisfaction and Loyalty

A.DATASET DESCRIPTION

This study utilizes a large-scale airline passenger satisfaction dataset comprising 129,880 post-flight survey responses collected from passengers traveling with US-based commercial airlines [23]. The dataset is publicly available on Kaggle: https://www.kaggle.com/datasets/teejmahal20/airline-passenger-satisfaction.

Each entry in the dataset corresponds to a completed flight and includes a combination of demographic variables, travel-related details, and subjective evaluations of service quality across various stages of the air travel experience. Passenger demographics include age, gender, frequent flyer membership status, and travel purpose (business or personal). Travel characteristics cover flight class, flight distance, and delay times. The core of the dataset consists of 14 service quality attributes, each rated by passengers on a standardized five-point Likert scale, ranging from 1 (very bad) to 5 (very good). These attributes span multiple service phases, including online booking, airport facilities, in-flight amenities, and onboard services. The target variable, passenger satisfaction, is originally recorded as a binary categorical label with two distinct values, “Satisfied” and “Neutral or Dissatisfied.”

The dataset was pre-partitioned into two subsets: a training set comprising 103,904 records (80%) and a test set comprising 25,976 records (20%). The class distribution in the training set is 57% “Neutral or Dissatisfied” and 43% “Satisfied,” while the test set maintains a similar proportion, with 56% “Neutral or Dissatisfied” and 44% “Satisfied.” This consistent stratification across splits ensures that model evaluation reflects realistic performance expectations on unseen data.

Although the dataset exhibits a moderate class imbalance, the distribution remains relatively balanced (57% vs. 43%). Therefore, no synthetic oversampling techniques such as SMOTE were applied. Instead, stratified train–test splitting was preserved to maintain the original class proportions across both subsets, ensuring fair and reliable model evaluation. In addition, performance was assessed using multiple complementary metrics, including precision, recall, F1-score, area under the receiver operating characteristic curve (ROC-AUC), and area under the precision-recall curve (PR-AUC), to avoid biased conclusions based solely on accuracy.

Overall, this dataset enables a comprehensive exploration of the determinants of passenger satisfaction, offering both behavioral and perceptual insights across multiple service dimensions. A detailed summary of the dataset’s attributes is presented in Table I.

Table I. Dataset attributes summary

B.PREPROCESSING

A comprehensive preprocessing strategy was implemented to ensure that the dataset was clean, consistent, and suitable for predicting passenger satisfaction. All transformations were embedded within a unified pipeline, guaranteeing that the same procedures were applied consistently across both training and testing phases, thereby eliminating the risk of data leakage.

Numerical attributes, including age, flight distance, departure delay, and arrival delay, were handled using a two-stage approach. Missing values were imputed with K-nearest neighbors (KNN) employing distance-based weighting, which leverages the similarity structure in the data to generate robust estimates. After imputation, the variables were standardized through z-score normalization to ensure that features with different scales contributed proportionally to model estimation. Next, service-related variables reflecting passenger perceptions of seat comfort, in-flight entertainment, food quality, cleanliness, and other service dimensions were retained as ordinal integers. Their missing values were filled with the most frequent response for each attribute, preserving the underlying distribution of passenger evaluations without distorting the rating scale.

Finally, categorical attributes such as gender, type of travel, and flight class were first handled using mode imputation to address missing values by replacing them with the most frequent category. Afterward, one-hot encoding was applied to transform each categorical feature into multiple binary variables, each representing the presence or absence of a specific category. This approach allowed the models to effectively capture nonlinear relationships and interaction effects between categories, while avoiding any bias or unintended assumptions that might result from using arbitrary numerical labels.

By integrating all preprocessing steps into a unified pipeline, the framework ensured both reproducibility and consistency across the entire workflow, while safeguarding the integrity of the evaluation process. This systematic data preparation established a solid foundation for training reliable machine learning models to predict passenger satisfaction. Once the dataset was fully cleaned and transformed, the focus shifted to selecting and developing the predictive models.

C.MODELING

Building a robust modeling framework for predicting passenger satisfaction requires more than selecting a single high-performing algorithm. Each machine learning algorithm embodies a distinctive inductive bias that reflects a set of assumptions about the underlying data distribution. This bias governs how the algorithm learns from examples, the types of relationships it is able to capture, and the patterns it emphasizes during prediction. Linear models are constrained to global linear relationships, neighborhood-based learners rely on local similarity, while tree ensembles are capable of representing nonlinear interactions and higher-order dependencies. Restricting the analysis to a single family of algorithms risks overlooking important structures within the data and leads to an incomplete understanding of the predictive space. To address this, the study employed a broad spectrum of algorithms ranging from simple baselines to state-of-the-art gradient boosting ensembles. This design ensured a balance between predictive performance and explanatory depth, providing both accurate forecasts and meaningful insights into the drivers of passenger satisfaction.

The objective of the modeling framework was not limited to maximizing classification accuracy but extended to deriving meaningful insights into why passengers report satisfaction or dissatisfaction. This dual perspective of prediction and interpretation guided the methodological design. Models with different representational capacities were complemented with multiple feature attribution techniques, including permutation importance, gain-based measures, and SHAP values [24]. The integration of these complementary approaches ensured that the insights were validated across diverse learning paradigms, moving the analysis beyond the question of which model predicts best toward the more actionable question of which factors consistently shape passenger satisfaction.

Crucially, feature attribution in this context extends beyond methodological validation. For airline practitioners, identifying the most influential service attributes provides a data-driven basis for prioritizing operational improvements. By highlighting whether factors such as seat comfort, boarding convenience, or delay management consistently drive satisfaction, the analysis translates predictive modeling into actionable guidance for resource allocation and service design. Importantly, the diversity of algorithms was not intended only to improve accuracy but also to strengthen interpretability for practitioners. In the airline industry, where strategic and operational decisions must be justified to stakeholders, explainable models play a critical role. By including both simple interpretable learners and more complex ensembles, the framework ensured that insights could be communicated in a form accessible to decision-makers, thereby enhancing the practical value of the findings.

To achieve these objectives, the study adopted a comprehensive model zoo that encompassed both simple and advanced algorithms. This zoo functioned as a form of evidence triangulation. Rather than relying on a single inductive bias, multiple algorithms were deployed to cross-validate explanatory signals. In this way, the analysis moved beyond optimizing performance metrics and provided converging evidence on which features consistently drive satisfaction, reinforcing the robustness of the conclusions.

At the baseline level, a Dummy Classifier (DM) with stratified outputs was included to establish a reference point against random guessing, ensuring that all subsequent models were evaluated relative to a meaningful lower bound [25]. Among linear models, Logistic Regression (LR) with L2 regularization was adopted as the canonical statistical classifier. It has long been considered a benchmark in supervised learning due to its ability to provide well-calibrated probability estimates and interpretable coefficients that reveal the global direction and strength of feature effects [26,27]. Its inclusion enabled the evaluation of whether more complex models achieved genuine performance gains or merely overfit the data.

To extend beyond the assumptions of LR, the framework incorporated a Linear Support Vector Classifier (LinearSVC) with probability calibration. Unlike LR, which minimizes log loss, LinearSVC maximizes the geometric margin between classes, making it more robust in high-dimensional spaces and potentially more effective for imbalanced decision boundaries [28]. Calibration was necessary to transform its decision scores into probability estimates, ensuring comparability with probabilistic models. In addition, a Stochastic Gradient Descent (SGD) classifier optimized with log-loss was evaluated, offering scalable training through stochastic optimization and flexible regularization for large, high-dimensional datasets [29].

Beyond the linear paradigm, the study employed KNN to capture local similarity structures. KNN is a nonparametric algorithm that predicts outcomes based on the proximity of an instance to its nearest neighbors in the feature space [30]. This makes it particularly sensitive to fine-grained variations in passenger demographics, travel conditions, and service ratings, allowing the framework to assess whether passenger satisfaction can be explained through localized behavioral and perceptual similarities.

To capture nonlinear dependencies in a more interpretable form, Decision Trees (DTs) were included. These models recursively partition the feature space into a hierarchy of rule-based decisions, providing a transparent view of how combinations of service attributes and passenger characteristics influence satisfaction [31]. While single trees are prone to overfitting, their interpretability makes them valuable for uncovering hierarchical predictor interactions that are readily accessible to industry practitioners. Building on this foundation, Random Forests (RFs) aggregated multiple DTs using bootstrap aggregation (bagging). By reducing the variance associated with individual trees, RF provided more stable and generalizable predictions while retaining a degree of interpretability. This ensemble approach offered a balance between accuracy and transparency, enabling the identification of consistently influential features across the models [32].

To further enhance predictive precision and address bias reduction, Gradient Boosting (GB) was incorporated. Unlike bagging, which builds trees in parallel, boosting constructs trees sequentially, with each iteration focusing on correcting the errors of its predecessors. This iterative refinement enables the modeling of complex nonlinear interactions and subtle dependencies [33]. Finally, three state-of-the-art boosting algorithms were integrated to capture the strengths of modern ensemble learning. XGBoost is widely recognized for its strong regularization mechanisms, scalability, and ability to handle sparse, high-dimensional data efficiently [34]. LightGBM was selected for its computational efficiency and scalability, achieved through histogram-based learning and leaf-wise tree growth, making it highly suitable for large-scale survey data [35]. CatBoost was incorporated for its robust handling of categorical variables and its use of ordered boosting, which mitigates prediction shift and enhances generalization [33]. Together, these implementations represented the frontier of gradient boosting and provided the strongest test of predictive performance in the study.

All models were encapsulated within a unified pipeline alongside preprocessing transformations, ensuring strict consistency between training and testing phases and eliminating the risk of data leakage. This systematic design established a modeling framework in which predictive accuracy and interpretability were equally prioritized, enabling the study not only to forecast passenger satisfaction with high fidelity but also to uncover the underlying drivers of passenger satisfaction to guide strategic decision-making and targeted service enhancements.

A.Evaluation Metrics

The evaluation of predictive models for passenger satisfaction cannot rely on a single metric, as different measures capture different aspects of model behavior. To ensure a rigorous and well-rounded assessment, the study employed a set of complementary metrics that jointly evaluate correctness, discrimination power, probability calibration, and diagnostic transparency.

At the most basic level, accuracy, as defined in Equation 1, quantifies the overall proportion of correctly classified passengers, offering an intuitive measure of performance [28]. However, accuracy alone does not reveal whether models favor one outcome over another. To provide a more detailed view, precision and recall were included. Precision, as defined in Equation 2, represents the proportion of passengers predicted as satisfied who were indeed satisfied [26], while recall, as defined in Equation 3, measures the proportion of all satisfied passengers that the model successfully identified [30]. Their harmonic mean, the F1-score, given in Equation 4 [33], balances these two dimensions, ensuring that improvements are not achieved at the expense of one over the other.

To assess the ability of models to discriminate between satisfied and dissatisfied passengers across varying thresholds, the framework incorporated the ROC-AUC. This metric reflects the probability that a randomly chosen satisfied passenger will be ranked higher than a dissatisfied one. Complementing this, the PR-AUC was used, as it directly evaluates the trade-off between precision and recall across thresholds, providing additional sensitivity to cases where both measures are equally critical [36]. Finally, confusion matrices were examined to provide a diagnostic perspective on model errors. By breaking predictions into true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN), confusion matrices reveal systematic tendencies in misclassification that may not be captured by aggregate metrics [34,37].

By integrating these complementary measures, the evaluation framework provides a multidimensional assessment of performance. This design ensures that the reported results reflect not only predictive accuracy but also the reliability, robustness, and interpretability of the models, thereby aligning with the study’s dual objective of prediction and explanation:

B.PREDICTIVE PERFORMANCE OF MODELS

The comparative evaluation revealed clear variation in performance across algorithms. Among linear classifiers, LR achieved an accuracy of 0.851, precision of 0.819, recall of 0.849, and an F1-score of 0.834, establishing a reliable but limited baseline. The calibrated LinearSVC performed slightly better in precision, 0.846, but attained a similar overall accuracy, 0.857, and F1-score, 0.835. The SGD classifier produced an accuracy of 0.854 and an F1-score of 0.832.

KNN and DT performed moderately. KNN obtained an accuracy of 0.925, precision of 0.943, and recall of 0.884, yielding an F1-score of 0.912. A single DT performed comparably, with an accuracy of 0.939, balanced precision of 0.930, and recall of 0.931, resulting in an F1-score of 0.931. While interpretable, both methods fell short of the top ensemble models.

Ensemble methods delivered markedly stronger results. RF achieved an accuracy of 0.957, precision of 0.965, recall of 0.937, and an F1-score of 0.951, highlighting the benefits of variance reduction through bagging. GB offered slightly lower overall accuracy, 0.935, but retained balanced precision, 0.935, and recall, 0.915, with an F1-score of 0.925.

The state-of-the-art boosting models outperformed all others. XGBoost achieved an accuracy of 0.959, precision of 0.966, recall of 0.941, and an F1-score of 0.953. CatBoost delivered almost identical results with an accuracy of 0.959, precision of 0.967, recall of 0.940, and an F1-score of 0.953. LightGBM emerged as the best-performing model, reaching the highest accuracy of 0.962 and F1-score of 0.956, supported by excellent precision of 0.969 and recall of 0.943.

Finally, as expected, the Dummy Classifier established the baseline at chance level, with an accuracy of 0.514 and an F1-score of 0.446, confirming the nontrivial predictive gains of all supervised approaches.

Table II presents a comparative summary of model performance. The results highlight the consistent superiority of gradient boosting ensembles, with LightGBM in particular achieving the most favorable balance across accuracy, precision, recall, and F1-score.

Table II. Comparative model performance

Macro-level performance measures capture the overall effectiveness of a model but often conceal the underlying structure of its errors. Confusion matrices make these error dynamics explicit by distinguishing true and false predictions for each class. In the context of passenger satisfaction, this diagnostic view is particularly valuable. Misclassifying dissatisfied passengers as satisfied risks masking service deficiencies, while the opposite error may exaggerate operational problems. Evaluating models through their confusion matrices, therefore, provides a complementary lens to summary metrics, offering both methodological validation and practical insight into how predictions succeed and where they fail.

To illustrate these dynamics, confusion matrices are presented in Tables III and IV for the weakest-performing model (DM) and the strongest-performing model (LightGBM), respectively. This contrast highlights the gap between random baseline behavior and the structured predictive power of advanced ensembles.

Table III. Confusion matrix—Dummy Classifier (baseline)

Table IV. Confusion matrix—LightGBM (best model)

The comparison shows that while the DM generates many false positives and false negatives, LightGBM achieves much higher correctness with very few false positives. This reduction is especially important, as it lowers the risk of overlooking dissatisfied passengers misclassified as satisfied.

C.FEATURE IMPORTANCE ANALYSIS

The objective of this study extends beyond building an accurate predictive model; it is equally concerned with uncovering the key drivers of passenger satisfaction. For airlines, understanding which service attributes most strongly influence satisfaction is essential for guiding strategic improvements, enhancing customer loyalty, and ultimately increasing profitability. To ensure that such insights are valid and actionable, they must be derived from highly accurate models—only when predictive performance is reliable can the factors highlighted as influential be trusted as true reflections of passenger priorities.

Accordingly, the analysis of feature importance was conducted using multiple complementary techniques, including permutation importance, gain-based measures from tree ensembles, and SHAP values. These approaches provide different perspectives on how features contribute to model predictions, offering both methodological robustness and practical interpretability.