Using PLS-SEM and XAI for causal-predictive services marketing research

Abstract

Purpose

This study aims to redefine approaches to metrics in service marketing by examining the utility of partial least squares – structural equation modeling (PLS-SEM) and eXplainable Artificial Intelligence (XAI) for assessing service quality, with a focus on the airline industry.

Design/methodology/approach

Using the Airline Passenger Satisfaction data set from Kaggle platform, this study applies PLS-SEM, facilitated by ADANCO software and XAI techniques, specifically using the SHapley Additive exPlanations TreeExplainer model. This study tests several hypotheses to validate the effectiveness of these methodological tools in identifying key determinants of service quality.

Findings

PLS-SEM analysis categorizes key variables into Delay, Airport Service and In-flight Service, whereas XAI techniques rank these variables based on their impact on service quality. This dual-framework provides businesses a detailed analytical approach customized to specific research needs.

Research limitations/implications

This study is constrained by the use of a single data set focused on the airline industry, which may limit generalizability. Future research should apply these methodologies across various sectors to enhance a broader applicability.

Practical implications

The analytical framework offered here equips businesses with the robust tools for a more rigorous and nuanced evaluation of service quality metrics, supporting informed strategic decision-making.

Social implications

By applying advanced analytics to refine service metrics, businesses can better meet and exceed customer expectations, ultimately elevating the societal standard of service delivery.

Originality/value

This study contributes to the ongoing discourse on artificial intelligence interpretability in business analytics, presenting an innovative methodological guide for applying PLS-SEM and/or XAI in service marketing research. This approach delivers actionable insights, not only in the airline sector but also across diverse business domains seeking to optimize service quality.

Citation

Goktas, P. and Dirsehan, T. (2024), "Using PLS-SEM and XAI for causal-predictive services marketing research", Journal of Services Marketing, Vol. ahead-of-print No. ahead-of-print. https://doi.org/10.1108/JSM-10-2023-0377

1. Introduction

Understanding the dynamics of business-related problems, and the associated explanation and prediction aspects, has become a critical concern for business researchers across a range of fields (Farooq et al., 2018; Leon and Leon, 2023). Although service quality assessment remains the area of primary focus (Parasuraman et al., 1985, 1988; Zeithaml et al., 1996), the wider application of problem-solving methodologies extends our research scope and contribution. This necessitates the use of robust tools and techniques for accurately depicting complex business situations, and includes the topics of service quality and customer satisfaction research.

The challenge in business research, particularly in service marketing, is to identify the most effective analytical methodology among a plethora of available options. Current methods such as Partial Least Squares – Structural Equation Modeling (PLS-SEM) and eXplainable Artificial Intelligence (XAI) have been widely applied and documented in the business research literature, such as this study conducted by Hair et al. (2021) on PLS and Arrieta et al. (2020) who discussed the applications of XAI. However, while PLS-SEM has limitations in terms of interpretability and robustness, leading to concerns about its balance between prediction and explanation (Cook and Forzani, 2023; Rönkkö and Evermann, 2013), XAI promises improved interpretability and reliability in data analysis. Yet, the question remains:

Q1.

When and how should they be strategically implemented in various service marketing scenarios?

This research addresses this fundamental query:

Q2.

What are the commonalities and differences between PLS-SEM and XAI, and when should each method be used?

Through this exploration, we use an illustrative case study focused on airline service quality assessment, elucidating the kind of outcomes each methodology can generate. Our aim is to guide researchers in making well-informed choices on methodological selection for their investigative endeavors.

In alignment with the focus on transforming service metrics, we not only explore PLS-SEM and XAI techniques and their applications in business research, especially in service quality assessment, but we also illustrate how these techniques can add to the explainability and interpretability in business applications. Furthermore, this research focuses on the incorporation of XAI into service marketing research, which can help businesses make better-informed decisions based on a thorough comprehension of the factors influencing customer satisfaction, even when a thorough SERVQUAL measurement is not practical. By understanding when and how to apply PLS and XAI techniques in various business problems, the business researcher can better apply these methods to design scenarios to achieve maximum business impact on service marketing practices.

2. Literature review and theoretical framework

2.1 Introduction to Partial Least Squares – Structural Equation Modeling

PLS-SEM developed from the pioneering work of Herman Wold in the 1970s, has been extensively used across various domains, including airline service quality (Sarstedt et al., 2022; Hair et al., 2019; Suki, 2014). This sophisticated multivariate statistical technique is engineered to maximize the explained variance of dependent latent constructs. It has been particularly favored in marketing and social sciences due to its capacity to handle complex models, small sample sizes and non-normal data distributions, which are common in these fields (Hair et al., 2011; Henseler et al., 2014; Chin et al., 2020). This section now includes a deeper exploration of foundational studies by Sarstedt et al. (2022) and others, linking service quality dimensions specifically explored within the airline sector to PLS-SEM capabilities. We discuss how these dimensions relate to customer satisfaction and decision-making in service contexts, providing a theoretical basis for the analytical choices in our study.

A distinctive feature of PLS-SEM is its ability to handle both reflective and formative measurement models, unlike covariance-based SEM (CB-SEM) which generally only accommodates reflective models. In reflective models, latent constructs cause the observed variables, whereas in formative models, observed variables construct the latent concept. This flexibility is crucial for exploratory research where theories may not be fully developed, allowing researchers to customize their analytical approaches to suit the data and theoretical structure at hand. Because its inception, PLS-SEM has undergone significant advancements aimed at enhancing its reliability and interpretability. Innovations such as the introduction of the heterotrait-monotrait ratio of correlations for assessing discriminant validity and bootstrap-based methods for robust statistical testing have addressed early criticisms regarding the method’s robustness (Henseler et al., 2016). For example, a fundamental PLS-SEM model is composed of several latent variables, represented as ovals and observed variables, shown as rectangles. Arrows demonstrate the direction of relationships between these variables. This configuration elucidates the model’s structural framework and the hypothesized interactions among variables, thereby offering a visual depiction of the model’s intricate dynamics and complexity.

PLS-SEM was chosen primarily for its robustness in handling complex models and its flexibility with data types, which is crucial in exploratory research where relationships among variables might not fully conform to the assumptions required by covariance-based SEM (CB-SEM) CB-SEM (Hair et al., 2011). PLS-SEM’s ability to handle formative and reflective measurement models offers a distinct advantage in exploring the intricate dimensions of service quality, where each dimension can potentially be a complex construct influenced by various observed variables. On the other hand, logistic regression (LaValley, 2008), while well-suited for binary outcomes, typically provides less insight into the complex structural relationships between latent constructs and observed variables. In our case, the use of single-item variables to represent different aspects of service quality necessitated a method that could effectively capture and model the latent constructs underlying these observed indicators, which PLS-SEM accommodates more naturally than logistic regression.

As for not opting for CB-SEM, the choice was influenced by the need for a method less sensitive to sample size and data distribution requirements. PLS-SEM is particularly advantageous in scenarios where the primary research aim is prediction and theory development, rather than theory testing. This aligns with our study’s objectives to both predict and explain factors influencing airline service quality, a domain where theoretical constructs are still being actively developed. Regarding the formative constructs, the configurational approach of PLS-SEM was deemed appropriate despite the simplicity of the model structure used in this study. This approach allows for a nuanced examination of how each service aspect, treated as a single item, contributes to the broader construct of service quality, adhering to the formative measurement model where the observed variables define the construct rather than merely reflecting it.

Despite these advancements, the interpretability and robustness of PLS-SEM remain points of contention. Concerns often focus on the optimum balance between prediction and explanation, with some critics arguing that PLS-SEM may prioritize prediction at the expense of providing deep insights into the underlying constructs it models (Cho et al., 2023; Cook and Forzani, 2023; Rönkkö et al., 2023). In addition, there is ongoing debate about the potential misapplication of PLS-SEM in confirmatory settings, where more traditional SEM approaches might be more appropriate (Rönkkö et al., 2016). Thus, PLS-SEM continues to be a dynamic and evolving tool in statistical modeling, offering a powerful alternative to traditional techniques, especially in situations where the assumptions of those techniques do not hold. Researchers are encouraged to apply PLS-SEM thoughtfully and stay informed about its latest developments and best practices to fully leverage its capabilities in their studies.

2.2 Introduction to eXplainable Artificial Intelligence

In response to the limitations of traditional modeling techniques such as PLS-SEM, the field of XAI has emerged as a significant method to enhance interpretability and reliability across various industries, including marketing (Zeng et al., 2021; Pranav and Gururaja, 2023). As AI increasingly influences decision-making processes, there is a growing need for AI outputs to be transparent and understandable, particularly in sectors such as healthcare, finance and airline service quality assessment, where decisions carry substantial consequences (Munoz and Laniado, 2021). XAI seeks to shift from the opaque “black box” approach of traditional AI to a “glass box” approach, enhancing stakeholder comprehension and trust in AI outputs. This is especially vital in areas such as personalized marketing and customer relationship management, where understanding AI’s decision-making processes can lead to better strategic decisions and increased user trust (Rai, 2019).

Recent advancements in XAI have focused on developing techniques that elucidate the AI decision-making process. Notable methods include SHapley Additive exPlanations (SHAP) and local interpretable model-agnostic explanations, which reveal how individual features impact predictions. These techniques play a critical role in not only enhancing model transparency but also in detecting and amending biases within AI models, thus promoting fairness and accuracy in AI-driven decisions (Cankaya et al., 2024). Moreover, XAI also has a significant impact in the education sector, where it helps to identify factors affecting student performance and completion rates, thereby enabling institutions to better support their students. By integrating XAI into educational data analysis, institutions can identify targeted interventions for at-risk students, potentially transforming educational outcomes and administrative strategies (Melo et al., 2022; Cankaya et al., 2024; Guleria and Sood, 2023).

The integration of XAI into data analysis within the marketing domain not only offers deeper insights into influential factors affecting service quality but also provides pointers for improvement. For example, in airline service quality assessment, XAI can explain why certain services or features are perceived better than others, guiding more effective resource allocation and service enhancements (Adadi and Berrada, 2018). Thus, XAI represents a transformative force at the intersection of technology and practical application, bridging the gap between advanced AI capabilities and the human requirement for comprehensible, actionable insights. The continuous development of XAI is crucial for the advancement of ethical AI practices and for fostering an environment where AI supports more informed, equitable and effective decision-making.

2.3 Synthesizing explanatory and predictive approaches

Combining explanation and prediction in business research allows for robust understanding and the development of accurate forecasts, serving as the foundation for better-informed business decisions (Shmueli et al., 2019). Figure 1 shows how PLS-SEM and XAI take different yet complementary paths to serve the dual purposes of explanation and prediction. Each method offers its own unique advantages, such as the predictive accuracy of PLS-SEM and the superior interpretability of XAI. To provide further clarification, we have included a comprehensive comparison of various SEM methodologies, including PLS-SEM, consistent PLS (Dijkstra and Henseler, 2015), Variance-based SEM (Reinartz et al., 2009), Covariance-based SEM (Hair et al., 2017), PLSpredict (Shmueli et al., 2019) and XAI techniques. This comparison, detailed in Supplementary Table 1, outlines each method’s main focus, developers, interface, model evaluation criteria, predictive accuracy, interpretability and applicability to complex models. This detailed comparison supports our methodological choice, highlighting PLS-SEM and XAI’s suitability for the complex, exploratory nature of our research into airline service quality.

Before testing our hypotheses, we conducted a brief comparison of two distinct methodologies and their respective tools used for PLS-SEM and XAI. For PLS-SEM, we primarily considered methodologies implemented via software such as SmartPLS (Ringle et al., 2015) and a modern approach to composite-based structural equation modeling (ADANCO, Hair et al., 2019; Henseler, 2020). In contrast, for XAI, we focused on methods that prioritize interpretability and explainability (Adadi and Berrada, 2018; Arrieta et al., 2020). A comparative overview of these methodologies is presented in Table 1.

In comparing PLS-SEM and XAI methods, it is important to consider the distinct objectives and strengths each methodology brings to model analysis. PLS-SEM, which can be executed through software such as SmartPLS (Ringle et al., 2015) or ADANCO (Hair et al., 2019; Henseler, 2020), focuses on comprehensive analysis and has been noted for its high predictive accuracy and applicability to complex models (Cho et al., 2023). Efforts to improve the predictive accuracy of PLS path models have recently been underscored by guidelines that emphasize the use of cross-validated predictive ability tests (Sharma et al., 2022). These guidelines build on earlier work by Sharma et al. (2021), which focused on prediction-oriented model selection in PLS-SEM. The aim is to further enhance the predictive accuracy of PLS path models through the integration of cross-validated predictions in model estimation. In addition, PLSpredict, as proposed by Shmueli et al. (2019), aims to improve the predictive accuracy of PLS path models through the integration of cross-validated predictions in model estimation. On the other hand, XAI methods, focusing on interpretability and explainability, are applied to make complex black-box machine learning (ML) models more comprehensible for businesses, enabling them to discern the mechanics that underpin the model’s predictions. Below, we highlight the advantages of XAI methods, such as interpretability, handling complexity, validation and robustness and flexibility in certain situations.

• Interpretability: XAI methods focus on providing clear explanations for the relationships and interactions among variables in complex models. This can help decision-makers better understand the underlying factors influencing customer satisfaction and lead to better-informed decisions for improving service quality.

• Handling complexity: XAI methods are effective in handling and analyzing those complex interactions among variables that might not be readily captured by traditional PLS-SEM techniques or PLSpredict.

• Validation and robustness: XAI methods can be used to validate the results from other analytical methods, such as PLS-SEM or PLSpredict, and provide a more reliable assessment of the factors influencing airline service quality.

• Flexibility: XAI methods can be applied to a wide range of black-box ML models, offering opportunities to compare and contrast the performance of various models in predicting customer satisfaction and understanding the factors driving that.

By understanding when and how to best apply either PLS or XAI techniques to various business problems, the business researcher can better use these methods to craft targeted scenarios that drive meaningful results (Cho et al., 2023; Gregor and Hevner, 2013; Henseler and Sarstedt, 2013; Stange et al., 2022). The findings of our study contribute to the ongoing discussion about the explainability and interpretability in artificial intelligence (AI) applications, and provide new insights for improving airline service quality assessment. Therefore, throughout this study, we aim to explore the potential to integrate XAI into airline service quality assessments to deepen the understanding of business challenges.

In this study, we investigated key questions regarding the application of either PLS-SEM or XAI techniques in assessing airline service quality:

• Does the XAI technique provide better insights into factors influencing customer satisfaction in the airline service quality context as compared to PLS-SEM?

• In which scenarios do either PLS-SEM or XAI techniques perform better when assessing airline service quality?

• Does the combination of explanation and prediction provide more comprehensive insights into factors influencing customer satisfaction in the airline service quality context?

• How effective is the proposed decision-making framework in guiding businesses to select between PLS-SEM and XAI techniques – or use them in combination – for airline service quality assessment?

3. Method

This section provides a detailed guide to implementing PLS-SEM using ADANCO (version 2.3.2, Hair et al., 2019; Henseler, 2020), and an XAI technique using SHAP TreeExplainer utility (version 0.41.0; Lundberg et al., 2020). These tools are specifically chosen for their effectiveness in addressing complex relationships among service quality variables and overall satisfaction, a core focus identified in Section 3.1. By elucidating the capabilities of PLS-SEM and XAI, this section not only demonstrates how these methodologies can be applied to intricate business research scenarios but also provides guidance on configuring a binary classification scenario, wherein data are segmented into one of two categories. This setup is crucial for investigating the dichotomous outcomes often encountered in service quality assessments.

3.1 Hypotheses development

Drawing upon the theoretical foundations outlined earlier, we develop our hypotheses to address specific research gaps identified in the literature. These hypotheses explore how various dimensions of airline service quality – such as in-flight services and delay management – affect customer satisfaction. We ground each hypothesis in established service marketing theories, which highlight the pivotal influence of diverse service attributes on customer perceptions and satisfaction. This theoretical anchoring ensures that our hypotheses are well-supported and relevant to the field of service marketing, drawing on seminal works by scholars like Parasuraman et al. (1985, 1988).

H1.

Delays, as integral elements of the service experience, have a substantial negative impact on customer satisfaction.

This hypothesis arises from the service failure literature, which consistently highlights that operational inefficiencies, such as delays, significantly deteriorate customer satisfaction (Zeithaml et al., 1996). Delays are particularly salient in the airline industry due to their direct and immediate impact on passengers’ travel plans (Suki, 2014). The negative association between delays and customer satisfaction is supported by the expectancy disconfirmation theory, which posits that failures to meet service expectations lead to dissatisfaction (Oliver, 1980).

H2.

Airport service quality is positively related to customer satisfaction.

This hypothesis is based on the servicescape model, which suggests that the physical environment and service process quality at the airport influence passengers’ overall service evaluation (Bitner, 1992). Positive interactions with airport services such as check-in, baggage handling and security checks are expected to enhance customer satisfaction by creating a seamless travel experience (Cronin and Taylor, 1992).

H3.

The quality of in-flight service, including factors such as seat comfort, in-flight entertainment and meals, positively affects customer satisfaction.

This hypothesis aligns with the theory of hedonic consumption, which argues that pleasure-related aspects of a service contribute significantly to overall satisfaction (Hirschman and Holbrook, 1982). In the context of airline services, in-flight amenities provide comfort and enjoyment, which are crucial for passenger satisfaction, especially in long-haul flights (Gilbert and Wong, 2003).

H4.

The combination of explanation and prediction provided by PLS-SEM and XAI techniques offers more comprehensive insights into factors influencing customer satisfaction compared to using either method alone.

This hypothesis is inspired by recent calls for more robust analytical approaches in service marketing research that integrate both predictive and explanatory powers (Shmueli et al., 2019). The integration of PLS-SEM and/or XAI is hypothesized to enhance the understanding of complex service quality dynamics by not only predicting outcomes but also explaining the relationships among variables (Rai, 2019).

3.2 Variable description

In the illustrative example, we opted to use openly available secondary data from Kaggle.net (Kaggle, 2024). This data source offers the advantage of being quick and cost-effective to obtain. However, it was also rigorously evaluated against various criteria such as methodology, accuracy, currency, objectivity, nature and reliability, as recommended by Malhotra (2020). The clear documentation and open comment feature of the data set assured us of its appropriateness for our analysis. It is important to note that the satisfaction variable in this data set is binary, classified as “neutral or dissatisfied” and “satisfied.” While it could include more categories if it had been collected directly by the researchers, the existing format remains highly suitable for PLS-SEM analysis due to the complexity of the exogenous variables involved. Furthermore, the application of XAI significantly enhances the interpretability of these binary outcomes. It achieves this by uncovering the underlying patterns and decision-making rules of the model, thereby providing deeper insights into the factors that most influence customer satisfaction and how various levels of service quality and customer interactions affect these satisfaction categories.

In addition, XAI’s ability to adeptly manage binary or categorical outcomes complements the structural modeling capabilities of PLS-SEM. This synergy allows for a more thorough examination of both direct and indirect relationships within the data. The primary goal of this study is to compare the analytical capabilities of PLS-SEM and XAI, rather than to delve into the antecedents of satisfaction. This comparative approach not only showcases how each method handles the binary outcome differently but also highlights the potential insights each method can provide, particularly in terms of model transparency and clarifying complex relationships among variables. The variables used in this study are listed in Table 2.

3.3 Data preprocessing

The process begins with data preprocessing, in which the data set must ideally provide a variety of data points pertinent to the research problem. Understanding a data set in depth is a fundamental step to verifying that the data is appropriately cleaned and transformed into a format suitable for subsequent analysis using either PLS-SEM or XAI techniques. Effective execution of preprocessing tasks involves importing the data set in a standard format (such as CSV files) into a suitable data analysis software tool or programming environment. For example, we can use Python on a Jupyter Notebook or a similar platform in the server environment that allows for efficient data analysis protocol. Missing values are handled either by using suitable imputation techniques – such as mean or median imputation – or by removing instances of missing data, based on the specific characteristics of the data set (Little and Rubin, 2019; van Buuren, 2018). Categorical variables are then converted into numerical variables, using techniques such as one-hot encoding and label encoding (Pedregosa et al., 2011). Such a procedure is crucial for preparing the data for analysis by statistical or ML algorithms that require numerical inputs. Depending on the specific requirements of the analysis techniques, the data may be standardized to ensure all variables are on a common scale. For example, this step can be achieved using methods such as MinMaxScaler (Patro and Sahu, 2015) or z-score standardization, also known as StandardScaler (Brownlee, 2020). Executing this data preprocessing step ultimately leads to more accurate and reliable results.

3.4 Partial least squares analysis, with ADANCO: Model specification, estimation and predictive performance assessment

Following data preprocessing, we move to the PLS analysis [e.g. using the ADANCO software (Hair et al., 2019; Henseler, 2020)]. The chosen variables are applied to establish relationships among them. These variables should reflect the business parameters that the researchers are interested in exploring. Please follow the following step-by-step guideline to conduct a PLS analysis:

• Categorize feature inputs: We start by organizing the data set’s feature inputs into categories we define. This step facilitates the organization of the PLS model by grouping associated variables.

• Specify the PLS model: We then define the PLS model, including latent variables, measurement items and structural relationships, based on the categorized feature inputs (Henseler and Sarstedt, 2013; Henseler et al., 2014). To do this, follow these steps:

  • – Identify latent variables (constructs) representing groups of related variables.

  • – Assign suitable measurement indicators to each latent variable.

  • – Define the structural relationships between the latent variables, indicating the hypothesized cause-and-effect relationships.

• Compute model parameters: Next, we calculate model parameters, including path coefficients and factor loadings (Henseler et al., 2016). To conduct this estimation, follow these steps:

  • – Import the preprocessed data (Section 3.3) into the PLS analysis tool.

  • – Specify the PLS model by entering the latent variables, measurement items and structural relationships defined in the previous step.

  • – Run the PLS estimation analysis to obtain model parameters, such as path coefficients (regression coefficients) and factor loadings (indicator weights).

• Assess the model’s performance: To evaluate the model’s accuracy and reliability, we validate the PLS models using techniques such as conducting collinearity diagnostics. In this stage, we can also compare the results with other relevant PLS models in assessing airline service quality.

3.5. eXplainable artificial intelligence analysis, with SHAP TreeExplainer: Model specification, estimation and evaluation of predictive performance

The subsequent stage involves the use of XAI techniques, specifically those in version 0.41.0 of SHAP TreeExplainer (Lundberg et al., 2020), for understanding the factors affecting particular business parameters and their relative importance. This approach identifies which individual indicator contributes the most to specific customer experience decisions. Please follow the following three steps to conduct an XAI analysis:

• Train a tree-based ML model: First, we train a black-box ML model, such as random forest (Breiman, 2001), using the same data and relationships as in the PLS model. Follow these steps:

  • – Split the preprocessed data set described in Section 3.3 into a training set and a test set, ensuring randomization and balance to prevent an imbalanced data set.

  • – Train the random forest model by using the training set and specify suitable hyperparameters.

  • – Assess the trained model’s performance on the test set.

• Evaluate prediction performance: This involves using the trained ML model to generate predictions for each observation in the data set. Follow these steps:

  • – Apply the trained model to the test set to obtain predicted values for the target variable (e.g. passenger satisfaction).

  • – Compare the predicted values to actual values to evaluate the model’s predictive performance by using relevant metrics such as accuracy, precision, recall and F1-score.

• Apply the selected XAI technique (e.g. SHAP TreeExplainer): Next, we use the SHAP TreeExplainer to the trained model to obtain global and local feature importance values and contribution scores for each observation (Lundberg et al., 2020). To perform this analysis, follow these steps:

  • – Import the SHAP library and initialize the TreeExplainer using the trained ML model.

  • – Compute SHAP values for each observation in the test set.

  • – Visualize the SHAP values to understand the importance of each feature and its contribution to the model’s predictions.

3.6 Implementing binary classification scenario in business research

This part provides an illustrative example of a binary classification scenario. The steps detailed here can be followed to distinguish between two distinct outcomes or states within your data set. A binary classification scenario is implemented, aiming to differentiate between “neutral or dissatisfied” and “satisfied” passenger intentions. Next, a tree-based ML model is trained (Section 3.5) and evaluated using the same data set and relationships as the PLS-SEM model (Section 3.4). The binary classification model’s performance is evaluated by applying the PLS-SEM and XAI analysis techniques to determine the most suitable approach for predicting passenger satisfaction. The relationship between normalized path coefficients or factor loadings and feature importance scores is analyzed either visually (e.g. scatter plot or bar chart) or by calculating correlation coefficients between feature sets. Evaluating the binary classification model’s performance by applying the PLS-SEM and XAI analysis techniques can provide insights into the most suitable approach for predicting a desired business outcome.

Overall, this guideline provides a comprehensive methodology for addressing business research problems using PLS-SEM and XAI techniques. Researchers can adjust this framework to their individual situations by adhering to the guidelines detailed in each subsection, thereby demonstrating the value of PLS-SEM and XAI in business research. This approach allows for effective evaluation and increased explainability of research processes.

3.7 Empirical exploration of Partial Least Squares – Structural Equation Modeling and eXplainable Artificial Intelligence for service quality assessment metrics

Kaggle, a well-recognized online platform, offers a wide range of data sets, competitions and tools in the fields of data science and ML disciplines (Kaggle, 2024). Within its data set environment, the Kaggle Airline Passenger Satisfaction data set is specifically of customer data collected from airlines, detailing customers’ experiences and satisfaction levels with various aspects of airline services (Kaggle, 2020). By analyzing this data set, we can obtain valuable insights into consumer satisfaction within the airline industry, thereby assisting in the formulation of strategies for improving airline service.

The objective of this research study is twofold. First, let us consider a hypothetical business researcher who intends to understand the structure of the data set and the various factors such as delay, airport service and in-flight service that influence overall customer satisfaction. This is accomplished using the PLS-SEM analysis, conducted by ADANCO (Hair et al., 2019; Henseler, 2020). The hypothetical researcher would also aim to explain the underlying relationships between these main factors and their respective subcomponents. For example, the relationship between airport service (the main concept) and its components such as check-in, baggage handling and online boarding as “ingredients.” The secondary objective of the business researcher is to predict passenger satisfaction levels. The researcher wishes to distinguish between passengers who are “neutral or dissatisfied” and those who are “satisfied.” This prediction task is mainly accomplished through the application of XAI techniques, particularly the SHAP TreeExplainer utility. It is important to note that, according to the literature (Parasuraman et al., 1985, 1988); although the public availability of the Kaggle data set allows the facilitation of a comprehensive SERVQUAL measurement, the role of PLS-SEM and XAI techniques in improving the comprehensibility of the assessment process is also emphasized. Applying the correct technique for service quality assessment enables businesses to make better-informed decisions based on a robust understanding of the drivers of customer satisfaction, even when a full SERVQUAL measurement may not be feasible.

In a service quality study, PLS-SEM and XAI techniques play crucial roles. The PLS-SEM method, through ADANCO software, aids in identifying latent variables and establishing relationships between chosen variables, thus addressing challenges such as multicollinearity and small sample sizes. On the other hand, the XAI method – specifically SHAP TreeExplainer – allows us to understand the output of tree-based ML classifiers, providing insights into the significance and contribution of individual indicators to decisions regarding customer experience. For the prediction task, the path coefficients from the PLS-SEM analysis, which indicate the strength and direction of relationships between variables, can serve as crucial information in the predictive analysis. Similarly, from the ML perspective, the accuracy, precision, recall and F1-score metrics can be used to evaluate the predictive performance of the applied (or trained) ML models in the various scenarios.

By applying either the PLS-SEM or XAI approach in the analysis of the Kaggle Airline Passenger Satisfaction data set, business researchers can create a comprehensive and interpretable model that facilitates effective assessment of airline service quality and highlights the key factors influencing passenger satisfaction. To access the Kaggle Airline Passenger Satisfaction data set (Kaggle, 2020), follow these steps:

• Sign in or create an account for the Kaggle platform (www.kaggle.com/) (Kaggle, 2024).

• Use the search bar to locate “Airline Passenger Satisfaction.”

• Find and click on the relevant training data set (e.g., “train.csv”).

In the Data Explorer section, click “Download” to obtain the data files in CSV format.

Having achieved a deep understanding of the Kaggle Airline Passenger Satisfaction data set, the business researcher can now progress to the actions described in Sections 3.3–3.6 to analyze the categorizes of Delay, Airport Service and In-flight Service. This analysis facilitates an in-depth assessment of airline service quality, differentiating between “neutral or dissatisfied” and “satisfied” passenger intentions in the data set.

Overall, our study used ADANCO version 2.3.2 to execute the PLS-SEM analysis on a data set of 25,976 observations. The detailed application of this approach is described in Section 3.4 of our study. Subsequently, for the utilization of the XAI technique, we implemented a series of procedural steps, outlined in the following five steps:

1. The “train_test_split” method was used to divide the Kaggle Airline Passenger data set into training and testing sets, with a test size of 30%. To ensure reproducibility in our results, a seed value (often referred to as “random state”) of 42 was set for the random number generator used in the splitting process.

2. Trained a random forest classifier using the following configuration: 100 estimators, no maximum depth, a minimum sample split of 2, the “Gini” criterion and a random state of 42.

3. K-fold cross-validation (Kohavi, 1995) with k = 10 was implemented due to the large size of the data set (approximately 103,905 data points).

4. Performance of the ML classifiers was evaluated using four classification metrics – accuracy, precision, recall and F1-score.

5. The SHAP TreeExplainer utility was applied to the trained Random Forest classifier to gain insights into the factors influencing airline service quality.

All experiments were conducted on a laptop computer (Dell Precision 5550: Intel® Core™ i7-10850H CPU @2.70 GHz 2.71 GHz RAM 32.00GB Windows 10 Pro).

4. Results

4.1 Partial Least Squares – Structural Equation Modeling analysis insights

Our experimental trials demonstrated that the PLS-SEM model accounts for 36.21% of the variance in passenger satisfaction, as indicated by an R-squared (R²) value of 0.362. Our analysis, based on the data set from Kaggle, which contained both training and test data, was performed directly without the typical PLS practice of training and testing due to the inability to import the test data (containing over 100,000 data points) into ADANCO software. Instead, the analysis was based on the available 20,000+ training data entries. We used 4,999 bootstraps; however, the necessity of bootstrapping for such a large data set in PLS remains uncertain.

In response to the concerns raised about the lack of a comparison benchmark and the limitations of using goodness of fit (GoF) indicators as the sole assessment of PLS-SEM model quality, we acknowledge the importance of focusing on predictive relevance. Unlike covariance-based SEM (CB-SEM), PLS-SEM does not optimize a global scalar function, and therefore does not use global GoF measures traditionally used in CB-SEM. Instead, PLS-SEM assesses the model’s predictive capabilities, focusing on the discrepancy between the observed or approximated values of the dependent variables and the values predicted by the model. This shift from GoF to predictive relevance is supported by the literature suggesting that PLS-SEM should primarily evaluate its predictive power, both in-sample and out-of-sample (Hair et al., 2019; Shmueli et al., 2019).

In our PLS-SEM analysis, variables such as “Inflight Wi-Fi service,” “Seat comfort,” “Inflight entertainment,” “On-board service,” “Legroom service,” “Baggage handling,” “Check-in service,” “Inflight service,” and “Cleanliness” are treated as independent predictors. These variables are not formative indicators of the satisfaction construct; instead, they act as antecedents of satisfaction, influencing the single-item dependent variable “satisfaction” independently. Each variable independently contributes to explaining the variance in passenger satisfaction. Our application of PLS involved Mode B calculations for the measurement of these variables, focusing on maximizing the explained variance of the dependent variable. As in our experimental trials, the model fit was found to be acceptable across all weighting schemes and discrepancies, as detailed in Table 3.

Table 4 illustrates the factor loadings (indicator weights) for the variables, offering significant insights for improving passenger satisfaction. Notably, Online Boarding has a high loading of 0.9046 and a weight of 0.8370, suggesting that it contributes significantly to customer satisfaction. In addition, Arrival Delay in Minutes shows a high loading of 0.9445 and a weight of 2.0005, indicating that arrival delay has a strong negative impact on customer satisfaction. Interestingly, Departure Delay in Minutes also has a high loading of 0.8043 but a negative weight of −1.1059. This suggests that departure delay also significantly influences customer satisfaction, but the relationship is negative, meaning that the longer the departure delay, the lower the customer satisfaction. The factor loadings for these aspects demonstrate the need for further investigation and service improvement to achieve passenger satisfaction levels.

Regarding the lack of robust comparison benchmarks and the inappropriate reliance on GoF indicators for evaluating the quality of the PLS-SEM model, we acknowledge the need for predictive relevance indicators. To enhance our model’s evaluation, we have conducted a collinearity diagnostics test, as the variance inflation factor (VIF), to ensure no bias affects our regression coefficients and rankings A VIF value exceeding 5 (some prefer a stricter threshold of 3) suggests significant collinearity that could bias the regression coefficients and the resulting importance rankings (Thompson et al., 2017). The VIF calculated for the loadings and weights array of the indicators in our PLS-SEM model are both approximately 1.48. For a detailed view of the VIF values across a broader set of variables, refer to Supplementary Table 2, which lists the VIF values for various service quality indicators and delay metrics. This low multicollinearity indicates a reliable estimation of the regression coefficients used in our model.

Despite advancements in the PLS-SEM model, we must consider that the relationships between variables in the Airline Passenger data set are nonlinear. This can result in an underestimation of the explained variance in satisfaction and limit the overall effectiveness of the analysis. In the structural model, Airport Service (path coefficient 0.3558) and In-flight Service (path coefficient 0.3116) emerged as significant factors for passenger satisfaction, whereas Delay (path coefficient −0.0406) was found to have a negative impact. The “Inflight Wi-Fi service,” “Seat comfort,” “Inflight entertainment,” “Onboard service,” “Legroom service,” “Baggage handling,” “Check-in service,” “Inflight service,” and “Cleanliness” indicators also contribute to passenger satisfaction, as indicated by their respective loading and weight estimates. The insights from this study could be valuable for airlines to improve their service quality and thereby increase customer satisfaction. However, we emphasize that a holistic understanding of passenger satisfaction would necessitate further investigations, especially concerning the PLS-SEM model’s ability to predict accurately using new data that it has not been trained on, often referred to as “holdout data.”

4.2 eXplainable Artificial Intelligence analysis insights

In the random forest classifier used for our study, the performance was marked by several key metrics. Of these, the Accuracy metric was particularly noteworthy in this case, with a value of 94.529%. The Precision, Recall and F1-score values were 94.458%, 93.982% and 94.189%, respectively. The application of k-fold cross-validation served to validate the model’s performance and mitigate overfitting, resulting in a more robust and reliable model for airline service quality assessment. We then used model-agnostic permutation importance and feature importance scores to evaluate the significance of input feature sets in the trained random forest classifier. In addition, the SHAP TreeExplainer utility provided relative feature importance, considering the balance between the ML model’s interpretability and efficiency on holdout data. In this context, SHAP values can be viewed as a latent representation of the Kaggle Airline Passenger Satisfaction data set, highlighting the most crucial parameters for understanding passenger satisfaction in the trained model.

In the binary classification scenario, where we aimed to distinguish between “neutral or dissatisfied” and “satisfied” passengers, the analysis identified the following crucial factors and their global feature importance ranking from the random forest classifier [Figure 2(a)]. Permutation feature importance is a method used to evaluate the significance of input features in an ML model. This is achieved by quantifying the reduction in the model’s performance when the values of a specific feature are randomly permutated. In the context of the Kaggle Airline Passenger Satisfaction data set, this technique helps us to understand the impact of each feature on the model’s ability to predict passenger satisfaction [Figure 2(b)]. The results indicate that factors, such as Departure Delay in Minutes and Seat Comfort, have the most significant impact on passenger satisfaction. On the other hand, the Online Boarding factor was found to have a lower impact in the permutation feature importance criteria than of the global feature importance score from the random forest classifier. It is crucial to acknowledge potential limitations and biases in the permutation feature importance analysis, such as multicollinearity among features or the inability to capture complex interactions between factors of the Airline Passenger Satisfaction data set.

To address the limitations and biases that may be present in the permutation feature or global feature importance analysis, we used the SHAP TreeExplainer utility to provide relative feature importance for each class assignment. One of the reasons for using SHAP is to handle the challenges associated with multicollinearity among features and the inability to capture complex interactions between features. Multicollinearity refers to the presence of high correlations between variables, which can lead to misleading interpretations of feature importance. The XAI-SHAP approach ensures that the contributions of correlated features are fairly distributed, thus minimizing the impact of multicollinearity on the feature importance rankings.

The SHAP summary plot, also known as the beeswarm summary plot, can provide both global and local important scores, allowing for a more detailed visual summary of each satisfaction class assignment in the Kaggle Airline Passenger data set, as shown in Figure 3. Departure Delay in Minutes, Seat Comfort, In-flight Wi-Fi Service, Onboard Service and Arrival Delay in Minutes emerged as the top five influential predictors in determining customer satisfaction [Figure 3(a)]. The analysis also revealed that shorter departure delays primarily contributed to higher customer satisfaction levels, and arrival delays in minutes demonstrated a trend toward increased satisfaction with lower levels of contribution [Figure 3(b)]. These insights emphasize the significance of these factors in shaping passengers' satisfaction in the airline industry. To interpret the SHAP values in the plots, we can use the following guidelines:

• The magnitude of the SHAP value represents the feature’s importance, e.g. how much it contributes to the model's prediction [Figure 3(a)].

• Positive SHAP values indicate that the feature has a positive effect on the model’s output, meaning that higher values of the feature increase the likelihood of the positive class [e.g. “satisfied” intention, Figure 3(b)].

Furthermore, through applying the XAI approach, the interactions of factors were analyzed by applying the SHAP dependence plot method, and we demonstrated a selection of these interactions to provide insights into their effects on passenger satisfaction (Figure 4). Based on the dependence plot analysis, we can interpret some representative interactions between the features in the dependence plot as described below.

• Seat Comfort and In-flight Wi-Fi Service:

If the seat comfort score is between 3 and 5, and the in-flight Wi-Fi service score is mostly higher than 2.5, then the SHAP value of seat comfort will be higher than zero. This indicates that in this range, both higher seat comfort and better in-flight Wi-Fi service have a positive impact on passenger satisfaction (assuming a higher SHAP value corresponds to higher satisfaction). To improve the customer experience and increase loyalty, airlines should invest in improving seat comfort and providing reliable and high-quality in-flight Wi-Fi services.

• Departure Delay and Arrival Delay:

Passengers are highly sensitive to even minor departure delays (e.g. level of one), which negatively impacts their satisfaction. However, when the departure delay exceeds a level of three, passengers' satisfaction becomes less sensitive to further delays, possibly because other possible factors become more crucial in determining satisfaction. This observation underlines the importance of on-time departures for airlines. It also suggests that in case of significant delays, airlines should focus on other aspects of service quality to reduce the impact on passenger satisfaction.

Through our analysis, to maximize passenger satisfaction and improve their sustainable competitive advantage, airlines should focus on improving: On-time performance, Seat comfort facility and In-flight Wi-Fi service. In addition, where delays are unavoidable, airlines should find a way to provide exceptional service in other areas, such as onboard service, legroom service and in-flight entertainment options, to help compensate for the negative impact of delays on passenger’ satisfaction [Figure 4(b)]. In essence, the application of SHAP augments our understanding of the crucial factors influencing Airline Passenger Satisfaction. By offering potential solutions for inherent constraints and biases linked with multicollinearity and complex feature interplays, SHAP outperforms traditional methods, such as permutation feature or global feature importance analysis, often used in random forest classifiers. This deeper insight facilitates a more reliable interpretation of the key determinants influencing customer satisfaction.

5. Discussion

5.1 Theoretical contributions

In the business marketing domain, the PLS-SEM method has been extensively used to explore the relationships between multiple variables in predicting and understanding the formation of customer satisfaction (Hair et al., 2011; Ringle et al., 2015; Hair et al., 2019; Shmueli et al., 2019; Henseler, 2020). However, concerns regarding the interpretability and robustness of the PLS-SEM method emerged, spurring debates about the balance between prediction and explainability (Rönkkö et al., 2016; Rönkkö et al., 2023). To address these concerns, this study aims to provide an overview of PLS-SEM and XAI techniques and to discuss their potential applications in business research, particularly in assessing airline service quality.

The Kaggle Airline Passenger Satisfaction data set provides a case study to illustrate the application of PLS-SEM analysis (using ADANCO) by grouping variables into Delay, Airport Service, and In-flight Service categories, and the application of an XAI technique (with SHAP TreeExplainer) to offer insights into the importance ranking of factors influencing perceived airline service quality. Based on the PLS-SEM analysis, as summarized in Tables 2 and 3, significant relationships were identified among the three latent variables (Delay, Airport Service, and In-flight Service) and customer satisfaction. The model’s goodness of fit was within acceptable limits, and the structural model showed that Airport Service and In-flight Service were positively associated with satisfaction, while Delay was negatively associated. In particular, certain indicators within these variables had significant loadings that impacted customer satisfaction. Reducing departure and arrival delays, improving the online boarding experience and offering a high-quality inflight Wi-Fi service could significantly augment overall passenger satisfaction. In addition, other notable indicators, such as Seat Comfort, Inflight Entertainment and Cleanliness, played substantial roles in increasing passenger satisfaction.

For the predictive task of the PLS side, the path coefficients from the PLS-SEM analysis, which indicate the strength and direction of relationships between variables, will also serve as crucial information in our predictive analysis. For example, variables that have high loadings and positive weights, such as Online Boarding and Arrival Delay in Minutes, will be expected to have a more significant influence on the prediction of passenger satisfaction. On the other hand, the XAI technique using the SHAP TreeExplainer utility provided a ranking of feature importance in predicting customer satisfaction, with Departure Delay in Minutes, Seat Comfort and In-flight Wi-Fi Service being the top three factors. The results of the SHAP analysis complemented the PLS-SEM findings, but also yielded additional insights into the specific aspects of in-flight service that have the most significant impact on customer satisfaction. Thus, the results of our PLS-SEM analysis and XAI technique using the SHAP TreeExplainer utility provided evidence to support the hypotheses.

To facilitate a clearer understanding and comparison between the XAI-SHAP and PLS-SEM methodologies, we have introduced a unified metric based on permutation importance – a technique widely used in XAI – to evaluate the impact of each predictor on model accuracy. For PLS-SEM, ensure that the feature importance is derived from measures such as path coefficients or loading values, which quantify the impact of each predictor on the outcome variable. In contrast, for XAI-SHAP, the feature importance should be based on SHAP values, which provide a measure of the impact of each feature on the model output. We also categorized key factors into three groups: Flight Service, Airport Service and Delay. This categorization involved recalculating the model's accuracy after individually removing or permuting each predictor, thereby directly quantifying each factor's contribution. The findings are displayed in Table 5, which provides a side-by-side comparison of the importance rankings from both methodologies, within the context of airline service quality, for a binary classification scenario.

From this quantitative comparison, it is evident that both methodologies generally agree on the relative importance of the majority of factors, such as Seat Comfort and Inflight Wi-Fi service, suggesting a consensus on these variables’ impact on passenger satisfaction. However, there are minor discrepancies, such as the positions of Leg Room Service and Onboard Service in the In-flight Service category and Baggage Handling and Check-in Service in the Airport Service category, which are ranked differently by the two methods. Moreover, the Delay category presents an interesting reversal of ranks. While PLS-SEM ranked Departure Delay in Minutes higher than Arrival Delay in Minutes, the XAI-SHAP analysis suggested the opposite. This analysis, underpinned by our unified metric, provides a robust framework for evaluating predictor importance, highlighting nuances that may be obscured when relying solely on one method. The side-by-side comparative analysis based on this unified metric offers valuable insights for service business researchers into passenger priorities, enabling targeted improvements to enhance customer satisfaction.

By synthesizing results from both methodologies under a common performance measure, we offer a comprehensive view of how the absence of specific predictors affects model performance across methods. This approach not only validates the predictive relevance of our models but also fosters a deeper understanding of methodological strengths and limitations. The discussion of these differences sheds light on the underlying assumptions, data handling capabilities and inherent strengths of each method, providing a clear justification for their use in specific contexts of airline service quality assessment. We also investigated several hypothesis questions, described in Section 3, concerning the use of PLS-SEM and XAI techniques in assessing airline service quality, and offered our interpretations regarding those hypotheses (Table 6).

5.2 Managerial implications

The decision-making framework presented in this study helps businesses align their research goals, assess complexity and data, guide correct technique selection and improve insights into factors influencing customer satisfaction, especially in the airline service quality context. PLS-SEM analysis grouped variables such as Delay, Airport Service and In-flight Service, each offering actionable indicators for enhancing passenger satisfaction. The XAI technique, in contrast to PLS-SEM, adds an extra layer of interpretability and reliability, thereby providing actionable insights for strategic decision-making. Thus, understanding these key metrics can help businesses focus on improving aspects that significantly affect customer satisfaction, such as reducing departure and arrival delays or improving in-flight services.

5.3 Limitations and future research

Our findings not only illuminate the strengths of both PLS-SEM and XAI but also demonstrate how these findings align with or challenge the theoretical expectations discussed in Section 3.1. We critically analyze how the variations observed in service quality dimensions impact customer satisfaction, thus contributing to theoretical and practical advancements in service marketing research. For instance, while PLS-SEM ranked Departure Delay in Minutes higher than Arrival Delay in Minutes, the XAI-SHAP analysis suggested opposite. Such variations highlight the value of using a multi-method approach to achieve a more comprehensive understanding that may be overlooked when relying on a single technique.

In addition, the decision to use a binary dependent variable in this study, though unconventional for PLS-SEM, was driven by the nature of the available secondary data. The binary coding of the satisfaction variable was not initially intended but necessitated by the data set’s format. This approach aligns with the flexibility of PLS-SEM in handling various data types, including binary outcomes, as a means of including categorical controls or moderators in models. However, as PLS-SEM relies on ordinary least squares (OLS) regression, using a binary variable as the ultimate dependent variable does indeed deviate from the ideal application of the method, as discussed by Hair et al. (2012).

Furthermore, while PLS-SEM is adept at analyzing secondary data with its flexibility in handling different data and measurement types (Hair et al., 2022), the use of a binary-coded construct as a dependent variable in an OLS framework does challenge the foundational assumptions of the methodology (Schuberth et al., 2018). Experts like Becker et al. (2023) and Ringle et al. (2023) suggest using binary data as control variables, moderators, or in grouping variables for PLS-SEM multigroup analysis, rather than as primary dependent variables. In this context, the use of XAI was intended to complement the PLS-SEM analysis by enhancing interpretability and addressing potential biases introduced by the binary nature of the data. Although our analyses demonstrate the operational feasibility of both methods under these conditions, they should not be interpreted as definitive comparisons of their efficacy. Instead, our aim was to showcase how XAI can serve as a valuable adjunct to PLS-SEM, particularly in scenarios involving non-ideal data types.

Future research could explore the integration of these methods for a more robust analysis, examining the origins of observed inconsistencies and developing strategies for their resolution. We investigated several hypotheses described in Section 3, concerning the use of PLS-SEM and XAI techniques in assessing airline service quality. However, a more comprehensive investigation across diverse business scenarios could also provide greater validity and possibly uncover nuances overlooked in this research.

6. Conclusion

Our study explored the use and comparison of PLS-SEM and XAI techniques in business research applications, specifically for the assessment of Airline Passenger Satisfaction. The four key takeaways from our research are set out below:

• PLS-SEM and XAI techniques each present distinct advantages and limitations within the business research field, such as in the context of airline service quality assessment. Importantly, we have enhanced the evaluation of these techniques by incorporating collinearity diagnostics and a unified performance metric for predictive accuracy, refining our understanding of each method’s effectiveness.

• The XAI approach through tools like the SHAP TreeExplainer, complements PLS-SEM by providing a deeper insight into the impact of each predictor, enhancing the interpretability of complex models. This integration aids in pinpointing critical factors that influence passenger satisfaction and their relative importance.

• The proposed framework equips businesses with a robust analytical tool for assessing and improving service quality, offering a detailed guide to navigating complex data landscapes and enhancing strategic decision-making.

• Further research is encouraged to validate our findings across different industries and contexts, expanding the applicability of our methodologies and exploring their potential in real-time customer satisfaction management.

As AI models become increasingly common in business decision-making, the transparency of their predictions and insights is crucial. Thus, this method not only clarifies complex analyses more understandable and clearer, but also furnishes airlines and other service providers with critical data to enhance their operations and ultimately boost customer satisfaction. By quantifying the relative importance of the key factors, businesses can prioritize interventions and allocate resources more effectively, resulting in increased customer satisfaction and loyalty.

Our PLS-SEM analysis emphasized the significant positive impact of In-flight Service and Airport Service factors on customer satisfaction over Delay, indicating a shift in customer priorities. Although delay minimization remains crucial, the quality of In-flight and Airport Services seems to be gaining increased importance in shaping customer satisfaction. This underestimated the need for airlines to pay closer attention to these areas for improving passenger satisfaction and overall service quality. This nuanced understanding prompts airlines to focus more on these service areas to improve customer experiences substantially.

In the XAI process, we used the SHAP TreeExplainer utility to interpret the output of our decision-tree ML models. The SHAP TreeExplainer utility is an explainable AI tool that interprets the predictions of complex models by assigning to each feature an importance value for a particular prediction. In this way, this tool assigns importance values to each feature for individual predictions, helping us to understand the contribution of each feature to the overall prediction in the scenario. We verified that the chosen variables consistently represented the constructs of interest, and that the relationships between constructs were meaningful and not merely due to random chance. This robust methodological approach underscores the complementarity of SHAP explanations in PLS context, ensuring the reliability and validity of our findings.

Our research demonstrates that XAI approaches, such as SHAP TreeExplainer, can successfully complement PLS analysis in identifying critical factors influencing passenger satisfaction and their importance ranking. Following design science principles from Gregor and Hevner (2013) and Stange et al. (2022), our study is structured to distinguish between the problem space (limitations of PLS) and the solution space (integrating XAI as a complementary approach). It is crucial to mention that, according to the literature, our available data does not allow for a complete SERVQUAL measurement, resulting in an arguably insufficient assessment (Parasuraman et al., 1985, 1988). Therefore, the primary aim of our study is not to provide a comprehensive measurement of SERVQUAL but to focus on the methodological contribution of increasing the explainability of the analysis using XAI. The guidelines presented in this study help businesses identify the most suitable method, depending on their objectives, data set, and context. By highlighting the benefits of integrating PLS and XAI in business research, we aim to contribute to the ongoing discussion about the importance of explainability and interpretability in service marketing research.

In conclusion, the evidence from our research substantiates each of the hypotheses set forth, underlining the viability of integrating PLS-SEM and XAI methodologies for a more nuanced evaluation of airline service quality. Our research not only achieves a rigorous comparative analysis of PLS-SEM and XAI techniques but also transcends this by presenting a structured framework for businesses and researchers to use these methods optimally. This contribution sets the stage for the development of AI-enhanced decision-support systems that deliver actionable, understandable insights. In doing so, we empower businesses to anticipate customer needs more effectively and elevate service quality. Therefore, our study serves a dual purpose: it offers both a nuanced comparison of analytical methodologies and a pragmatic roadmap for their successful implementation in the business context.

Figures

Figure 1

Trade-off between model interpretability and accuracy

Figure 2

The feature importance of the ML model

Figure 3

Relative feature importance using SHAP explanations

Figure 4

The SHAP dependence plot highlights the interactions between key factors influencing Airline Passenger Satisfaction

Table 1

Comparison of PLS-SEM and XAI methods for business-related problems

Feature	PLS-SEM	XAI
Developers	Various authors (Ringle et al., 2015; Hair et al., 2019; Henseler, 2020)	Multiple researchers and developers (Adadi and Berrada, 2018)
Main focus	Comprehensive PLS-SEM analysis	Interpretability and explainability in AI
Interface	User-friendly (e.g. ADANCO, SmartPLS)	Specific to the XAI technique/tool
Model evaluation criteria	Incorporates advanced features and algorithms; includes goodness of fit measures	Varies depending on the XAI method
Predictive accuracy	Good, but may vary depend on the method and model	Good, but dependent on the particular machine learning model
Interpretability and explainability	Fair	Excellent
Applicability to complex models	Good	Excellent
Prediction, explanation	Balanced emphasis	Explanatory emphasis

Source: This information was compiled by the authors from a comprehensive review of related literature, supplemented with ratings based on the author’s expertise

Table 2

Description of the variables

Variable	Description
Inflight Wi-Fi service	Satisfaction level of the inflight Wi-Fi service (0: Not applicable; 1–5)
Food and drink	Satisfaction level of food and drink (1–5)
Online boarding	Satisfaction level of online boarding (1–5)
Seat comfort	Satisfaction level of seat comfort (1–5)
Inflight entertainment	Satisfaction level of inflight entertainment (1–5)
Onboard service	Satisfaction level of onboard service (1–5)
Legroom service	Satisfaction level of leg room service (1–5)
Baggage handling	Satisfaction level of baggage handling (1–5)
Check-in service	Satisfaction level of check-in service (1–5)
Inflight service	Satisfaction level of inflight service (1–5)
Cleanliness	Satisfaction level of cleanliness (1–5)
Departure delay in minutes	Minutes delayed at departure
Arrival delay in minutes	Minutes delayed at arrival
Satisfaction	Airline satisfaction level (satisfaction, neutral or dissatisfaction)

Source: Kaggle (2020), “Airline Passenger Satisfaction data set. kaggle”. Retrieved from www.kaggle.com/datasets/teejmahal20/airline-passenger-satisfaction

Table 3

PLS-SEM composite analysis results

Weighting scheme	Discrepancy	Value	HI95	HI99	Result of assessment	Conclusion
Saturated model	SRMR	0.0804	0.0048	0.0054	Pass	Model has acceptable fit
	dULS	0.6790	0.0025	0.0031	Pass	Model has acceptable fit
	dG	0.1605	0.0006	0.0007	Pass	Model has acceptable fit

Table 4

Weight and loading estimates of the indicators

Indicator	Loading	Weight
Inflight Wi-Fi service	0.5295	0.3567
Food and drink	0.4096	−0.0915
Online boarding	0.9046	0.8370
Seat comfort	0.6540	0.3860
Inflight entertainment	0.7534	0.1401
Onboard service	0.6072	0.3092
Legroom service	0.5848	0.3034
Baggage handling	0.4617	0.3456
Check-in service	0.4417	0.1886
Inflight service	0.4638	0.0433
Cleanliness	0.5937	0.1774
Departure delay in minutes	0.8043	−1.1059
Arrival delay in minutes	0.9445	2.0005
Satisfaction	1.0000	1.0000

Table 5

Comparative importance ranking of XAI-SHAP and PLS-SEM methods for in-flight service quality, airport service and delay variables, in a binary classification scenario

Factors	PLS-SEM	XAI-SHAP
Inflight service
Seat comfort	1	1
Inflight Wi-Fi service	2	2
Legroom service	4	3
Onboard service	3	4
Inflight entertainment	5	5
Cleanliness	7	6
Food and drink	6	7
Airport service
Online boarding	1	1
Baggage handling	3	2
Check-in service	2	3
Delay
Arrival delay in minutes	2	1
Departure delay in minutes	1	2

Table 6

Comparative analysis of PLS-SEM and XAI techniques applied to various scenarios/factors in business research problems

Factor	PLS advantages	XAI advantages
Better insights	- Estimates complex correlations between latent variables - Effective in formative measurement scenarios	- Improved explainability and interpretability - Addresses PLS limitations - Improved accuracy - Emphasis on prediction
Small sample sizes and non-normal data distribution	- Handles small sample sizes and non-normal data distributions more effectively than traditional covariance-based techniques
Complex relationships among latent variables	- Effectively estimate complex correlations between latent variables and provide insights into their overall impact on customer satisfaction	- Helps businesses understand co-interactions among variables and identify key drivers of customer satisfaction - Uncovers hidden patterns and intricate relationships that are not apparent using traditional analytical techniques
Formative measurement scenarios	- Provides accurate estimates in formative measurement scenarios, where indicators cause the latent variables rather than being caused by them
High interpretability requirement		- Provides additional insights into the importance and contribution of specific factors - Offers clear explanations of factors driving customer satisfaction
Validation and robustness concerns		- Validates results from other analytical methods. - Provides a more reliable assessment of factors influencing airline service quality - Increases confidence in analysis results
Combination for more comprehensive insights	Complementary strengths when applying XAI techniques. - Contributes to improved decision-making - Improves analysis of complex relationships and interactions among variables	Complementary strengths when applying PLS techniques - Contributes to improved decision-making - Improves analysis of complex relationships and interactions among variables

Source: This information was compiled by the authors from a comprehensive review of related literature, supplemented with rating based on the author’s expertise

Supplementary material

The supplementary material for this article can be found online.