Explainable AI and Fuzzy Linguistic Interpretation for Enhanced Transparency in Public Procurement: Analyzing EU Tender Awards

Abstract

Despite the ideal of a unified Single Market, a powerful “home bias” pervades EU public procurement, hinting at unseen barriers that conventional analysis fails to capture. This study introduces an interpretable AI framework to investigate these dynamics, pairing a LightGBM model with SHapley Additive exPlanations (SHAP) to examine the vast Tenders Electronic Daily (TED) database (2018–2023). Concretely, we propose a fuzzy linguistic layer that translates SHAP’s complex quantitative outputs into intuitive, human-readable terms. Our model effectively distinguishes local from non-local awards (AUC ≈ 0.855), revealing that while high-value contracts expectedly attract broader competition, the most potent predictors are a country’s own history of local awards and structural factors like the buyer’s type and location. This points not to isolated incidents, but, rather, to deep-seated patterns shaping market fairness. Our combined XAI-Fuzzy approach offers a new instrument for transparent governance, enabling policymakers to diagnose market realities and forge a more genuinely open and equitable European public square.

1. Introduction

Public procurement represents a cornerstone of the EU economy, with public authorities annually spending approximately 14% of the collective GDP—over EUR 2 trillion—on essential goods, works, and services [1]. This expenditure is governed by the core principles of the Single Market: transparency, equal treatment, non-discrimination, and open competition, all intended to foster robust cross-border activity [2]. However, despite this clear policy framework, a powerful “home bias” pervades the market. Direct cross-border procurement remains stubbornly low, with most contracts awarded to domestic suppliers, revealing unseen barriers that conventional analysis fails to capture [3]. Diagnosing the reasons for this gap between policy and reality—from administrative hurdles to implicit national preferences—is a paramount challenge for achieving the EU’s goal of smart, efficient, and transparent governance [4]. The vast and complex data within the Tenders Electronic Daily (TED) platform [5] offers a rich opportunity, but its scale and intricacy demand more powerful analytical tools than traditional statistics can provide.

To bridge this analytical gap, this study introduces an interpretable AI framework to dissect the drivers of public procurement outcomes. The “black box” nature of many advanced machine learning models hinders their adoption in public governance, where accountability and the ability to understand why a decision was made are non-negotiable [6]. We address this by pairing a high-performance LightGBM model with SHapley Additive exPlanations (SHAP), a leading Explainable AI (XAI) method that attributes model predictions to the underlying features in a fair and transparent manner [7]. Going a step further, we introduce a novel fuzzy linguistic interpretation layer that translates SHAP’s complex quantitative outputs into intuitive, human-readable terms. This innovation aims to make the intricate dynamics of market behavior accessible to non-technical stakeholders, such as policymakers and oversight bodies. By applying this framework to a comprehensive dataset of EU tenders from 2018 to 2023, we seek to answer the following research questions:

RQ1:

What are the primary characteristics of public tenders (e.g., value, competition level, buyer type) that drive the likelihood of a contract being awarded to a local versus a non-local supplier?

RQ2:

How can SHAP values, further clarified through a fuzzy linguistic framework, illuminate the specific ways these features influence outcomes, revealing patterns indicative of market integration levels or systemic biases?

RQ3:

What practical insights for smart governance and evidence-based policymaking can be derived from this interpretable machine learning framework?

This study’s contribution is threefold. Empirically, we provide data-driven evidence identifying the key factors that shape the EU’s procurement landscape. Methodologically, we demonstrate a novel, integrated XAI-Fuzzy framework designed for enhanced transparency and accessibility in complex policy analysis. Practically, we offer a powerful diagnostic tool that enables policymakers to better understand market realities, monitor fairness, and forge a more genuinely open and equitable European public square. The remainder of this paper details our methodology in Section 3, presents the model performance and explanatory findings in Section 4, discusses their implications in Section 5, and offers concluding remarks in Section 6.

2. Literature Review

This section synthesizes the literature on EU public procurement, the application of AI and XAI in its analysis, and the methodological foundations for enhancing interpretability, thereby contextualizing our study’s contribution.

2.1. Regulation and Market Challenges in EU Public Procurement

Public procurement is a cornerstone of the EU economy, vital for delivering public services and stimulating growth. Its governance rests on core principles of transparency, equal treatment, and open competition, enshrined in EU Treaties and a series of evolving directives [8,9,10]. The Tenders Electronic Daily (TED) platform serves as the central repository for procurement notices and the primary data source for analysis, although its utility is conditioned by known data quality issues that necessitate careful pre-processing [11].

Despite this robust regulatory framework, a significant gap persists between policy ideals and market realities. Direct cross-border procurement shows persistence in remaining low—often below 5% by value—indicating powerful “border effects” rooted in administrative complexities, national preferences, and information barriers [3,12,13]. These challenges are often magnified for small and medium-sized enterprises (SMEs), which face disproportionate hurdles related to procedural complexity and financial constraints, despite policies designed to support their participation [10,14]. Compounding these issues, overall competition has seen a decline, marked by an increase in single-bid procedures and direct awards [1].

This complex landscape is further shaped by policy initiatives like the 2014 directives, which promoted the “Most Economically Advantageous Tender” (MEAT) criterion. This principle advocates for a holistic value assessment over simply the lowest price, encouraging the inclusion of quality, environmental, social, and innovation aspects [2]. However, in practice, a large number of contracts continue to be awarded based primarily on price [10,15]. This implementation gap not only hinders the achievement of strategic goals like innovation or Green Public Procurement [16,17,18], but also adds another layer of complexity to an already challenging environment. The intricate interplay of these legal, economic, and procedural factors suggests that subtle, systemic forces are at play, requiring analytical methods that can look beyond conventional statistics to uncover them.

2.2. From AI to XAI in Public Procurement

In response to this complexity, researchers and practitioners are increasingly turning to artificial intelligence (AI) and machine learning (ML) to extract deeper insights from large-scale procurement data. These technologies offer transformative potential for improving efficiency, generating cost savings, and enhancing decision-making [19,20]. Applications are diverse, ranging from using ML models on tender data to detect fraud and collusion [21] to leveraging natural language processing (NLP) for bias detection in tender documents [22]. Furthermore, ML is being applied to predictive analytics for tasks like award prediction, supplier risk assessment, and demand forecasting [23], while NLP also facilitates the automatic classification and analysis of tender texts [24].

However, the growing power of these models is often accompanied by a critical drawback: many advanced ML systems operate as “black boxes”, with their opaque internal logic hindering their adoption in high-stakes public domains where accountability, fairness, and transparency are non-negotiable [6,25,26]. This lack of transparency erodes public trust and critically complicates the processes of auditing models for errors or correcting for systemic biases [27]. In recognition of this challenge, international bodies like the OECD have made explainability a core tenet of their principles for trustworthy AI [28].

Explainable AI (XAI) has emerged as the critical field dedicated to developing methods that make AI model outputs interpretable to humans [29]. Among the various techniques, SHapley Additive exPlanations (SHAP) has become a leading approach due to its solid theoretic foundation for fairly attributing a model’s prediction to its input features [7]. For tree-based models such as the LightGBM used in this study, the TreeSHAP algorithm offers an efficient method to compute exact and consistent explanation values, a notable advantage over local approximation methods like LIME, which can sometimes suffer from instability [30,31]. The imperative for such explainability is not merely technical; it is fundamental for public policy, regulatory compliance, and democratic accountability in an era of algorithmic governance [27,32,33]. This is further underscored by emerging regulations like the EU AI Act, which imposes stringent transparency and oversight requirements for high-risk AI systems, implicitly necessitating XAI capabilities [34,35].

2.3. Fuzzy-Linguistic Interpretation of XAI Results

Applying this XAI paradigm to EU public procurement offers a powerful new method for assessing market integration and diagnosing potential biases. Traditional analyses have long identified persistent barriers to a fully integrated market, including regulatory divergence and information asymmetries [36], but have struggled to quantify the precise influence of specific tender characteristics. XAI moves beyond correlation to attribute outcomes directly to contributing factors, providing a lens to reveal implicit barriers or systemic biases that conventional methods may miss [26,37]. For instance, if certain buyer categories or procedural choices consistently favor domestic suppliers after controlling for economic variables, XAI can highlight and quantify these nuanced associations [22]. Such insights are vital for upholding the EU’s core principles of fairness and non-discrimination, offering a robust method to scrutinize whether automated systems maintain these values and mitigate biases stemming from data or models [38,39].

While XAI provides invaluable quantitative explanations, the numerical outputs of methods like SHAP can still pose an interpretive challenge for non-technical stakeholders such as policymakers and auditors. To bridge this final gap between technical precision and actionable insight, our study introduces a novel interpretation layer based on fuzzy set theory. Public procurement, especially under the MEAT framework, is inherently a multi-attribute decision-making (MADM) or multi-criteria decision analysis (MCDA) problem, where authorities must evaluate tenders against multiple, often conflicting, criteria [2,40,41]. While formal MADM methods like AHP and TOPSIS provide structured frameworks [42,43], their outputs can also lack intuitive transparency.

Fuzzy set theory, introduced by Zadeh [44], provides a mathematical framework for handling the uncertainty and vagueness inherent in human language. By creating “linguistic variables”—variables whose values are words like “Low”, “Medium”, or “High”—it becomes possible to map complex numerical outputs to intuitive, qualitative descriptors [45]. This approach has found success in various decision-making contexts, including MADM, by translating quantitative data into human-understandable terms [46,47]. By defining fuzzy sets over the range of SHAP values, we can categorize their impact using accessible labels like “Strong Positive Influence” or “Slight Negative Influence”. The choice of simple triangular membership functions for this task is common due to their interpretability and ease of implementation [48]. This layer of fuzzy linguistic interpretation, built upon the rigorous foundation of SHAP, aims to make the complex dynamics of public procurement more accessible, aligning with the principles of smart governance where clarity for a diverse audience is paramount.

2.4. Synthesis and Contribution

The literature demonstrates that while AI and ML are increasingly applied to public procurement, a primary focus on explanation for pan-European policy insight remains a developing area. Existing research has established the value of ML for predictive tasks and has begun applying XAI in specific national contexts. For instance, studies have used XGBoost and XAI to analyze contract variations in Italy [49] or developed ML systems to predict tender complaints [50]. Other work has focused on automating tender classification using large language models [51] or conducting structural analyses of market barriers for specific firm types [52]. To situate our study within this landscape, a comparative overview is presented in

Table 1. Comparative overview of recent AI/ML applications in public procurement analysis.

Study	Methodology	Data Scope	Primary Objective	Distinguishing Focus
Current Work (2025)	LightGBM, SHAP, Fuzzy Linguistic Interpretation	Pan-European (TED)	Explain market integration drivers	Pan-European explanatory analysis of winner nationality using a fuzzy-enhanced XAI framework.
Fatima, Nai, & Meo (2025) [49]	XGBoost, XAI	National (Italy)	Predict contract variations	XAI application for fairness analysis of contract variations within a single country’s procurement system.
Minghini et al. (2024) [51]	GPT-3.5, Support Vector Machine (SVM)	Pan-European (TED)	Classify tenders by domain	Automation of tender classification in a specific technological sector using ML on TED data.
Nai et al. (2023) [50]	ML Classifiers, NLP, Recommender System	National (Italy)	Predict tender complaints	Predictive analysis of an administrative outcome (complaints) and development of a functional tool for bidders.
Westermann et al. (2025) [52]	Quantitative Analysis, Firm Classification	Pan-European (TED), enriched with Orbis	Analyze SME/startup barriers in ICT	Structural analysis of market access barriers for SMEs and startups in the European ICT sector.

.

The existing body of work provides a strong foundation, yet a research gap persists in applying a dedicated XAI framework to the comprehensive pan-European TED dataset to explain the systemic drivers of market integration, rather than focusing on national-level prediction or classification. Furthermore, a significant opportunity exists to improve the accessibility of XAI findings for the non-technical stakeholders who shape policy. This study addresses these gaps by not only using an interpretable model to explain the drivers of local versus non-local awards across the EU Single Market but also by introducing a novel fuzzy linguistic layer designed to make these complex findings transparent, actionable, and policy-relevant.

3. Materials and Methods

This section outlines data sources, processing, modeling, and interpretability techniques for analyzing drivers of local vs. non-local contract awards.

3.1. Data Source and Cohort Definition

Data is from the Tenders Electronic Daily (TED) database via EU Open Data Portal [53], using Contract Award Notices (CANs) from 2018 to 2023. Cancelled tenders (25 notices) were removed. The target variable, is_local_winner, was derived by comparing winner (WIN_COUNTRY_CODE) and authority (ISO_COUNTRY_CODE) country codes (1 = local, 0 = non-local). Tenders with missing country codes necessary for this derivation were excluded. We acknowledge TED data limitations like missing fields and potential reporting biases. The final dataset comprised 4,274,324 notices, with 90.29% local and 9.71% non-local winners.

3.2. Feature Engineering

To prevent data leakage, information about the winning entity (e.g., WIN_NAME, WIN_COUNTRY_CODE) was removed prior to feature engineering related to the winner. We generated 27 new features based on contract value, competition levels, historical country performance, and other characteristics. These were combined with 10 original categorical variables to form an initial candidate pool of 37 features. Key features influencing award outcomes were extracted or derived, including:

• Numerical Features:

• AWARD_VALUE_EURO_FIN_1_log: Log-transformed final contract value in Euros.
• NUMBER_OFFERS_log: Log-transformed number of offers.
• historical_local_win_rate: Country-level historical rate of local wins, calculated using time-aware features from DT_DISPATCH.

• Categorical Features:

• ISO_COUNTRY_CODE: Buyer country code.
• MAIN_ACTIVITY: Main activity sector of the contracting authority.
• CAE_TYPE: Type of contracting authority.
• TYPE_OF_CONTRACT: Type of contract (e.g., works, services, supplies).
• no_competition: Binary indicator for tenders receiving only one offer.

Categorical features were label encoded for compatibility with LightGBM (e.g., MAIN_ACTIVITY had 108 unique values, ISO_COUNTRY_CODE had 33). Choices were guided by domain knowledge, data availability, and initial feature importance screening. After an initial model run to assess feature importance, 26 features were selected for the final tuned model. The full list of all features considered for the model is detailed in

Table A1. Full list of candidate features for the predictive model.

Feature Name	Description	Type
Value-Based Features
AWARD_VALUE_EURO_FIN_1_log	Log-transform of the final contract value.	Numerical
AWARD_VALUE_EURO_FIN_1_sqrt	Square-root transform of the final contract value.	Numerical
AWARD_VALUE_EURO_FIN_1_quartile	Quartile bin of the final contract value.	Categorical
Competition-Based Features
NUMBER_OFFERS_log	Log-transform of the number of offers received.	Numerical
competition_level	Categorical bin for number of offers (1, 2–3, 4–5, etc.).	Categorical
no_competition	Binary: 1 if the tender received exactly one offer.	Binary
high_competition	Binary: 1 if the tender received 5 or more offers.	Binary
Historical and Country-Level Features
historical_local_win_rate	The tender country’s historical rate of awarding to local winners.	Numerical
country_tender_count	Total number of tenders from the buyer’s country in the dataset.	Numerical
tender_in_eu15	Binary: 1 if buyer country is in the EU15 group.	Binary
tender_in_eastern_eu	Binary: 1 if buyer country is in the Eastern EU group.	Binary
Contract and Procedure Characteristics (Original)
ISO_COUNTRY_CODE	The 2-letter code of the contracting authority’s country.	Categorical
CAE_TYPE	The type of the Contracting Authority (e.g., National, Regional).	Categorical
MAIN_ACTIVITY	The main sector of activity for the contracting authority.	Categorical
TYPE_OF_CONTRACT	The primary type of the contract (e.g., Services, Works).	Categorical
TOP_TYPE	The type of procedure used (e.g., Open, Restricted).	Categorical
CRIT_CODE	Code for the award criteria (e.g., Price only, MEAT).	Categorical
B_GPA	Binary: Whether the tender is covered by the GPA.	Categorical
B_SUBCONTRACTED	Binary: Whether subcontracting is foreseen.	Categorical
Complexity and Temporal Features (Engineered)
has_eu_funds	Binary: 1 if contract is financed by EU funds.	Binary
is_framework	Binary: 1 if the tender is part of a framework agreement.	Binary
complexity_score	A composite score based on value, duration, funds, etc.	Numerical
high_complexity	Binary: 1 if complexity_score is high.	Binary
DT_DISPATCH_year	The year the tender notice was dispatched.	Categorical
DT_DISPATCH_quarter	The quarter the tender notice was dispatched.	Categorical
DT_DISPATCH_month	The month the tender notice was dispatched.	Categorical

in the Appendix C.

3.3. Predictive Modeling

LightGBM [54] was used for binary classification (is_local_winner) due to its efficiency and performance with large datasets and categorical features. Class imbalance was addressed using scale_pos_weight = 0.11 (ratio of non-local to local instances in the training set). Data was split into 70% train (2,992,026 instances)/30% test (1,282,298 instances) using stratification and random_state = 42.

Hyperparameter tuning was performed using grid search with 3-fold cross-validation, achieving a best cross-validation AUC score of 0.8534. The best parameters found were n_estimators = 500, learning_rate = 0.1, max_depth = 8, num_leaves = 63, colsample_bytree = 0.8, and subsample = 0.8. The model was trained with objective = ‘binary’ and metric = ‘auc’, employing early stopping with 30 rounds on a validation set during the initial feature selection phase, though the final model used the fixed 500 estimators from grid search.

3.4. Explainability Method (SHAP)

SHAP interpreted the LightGBM model. SHAP, from game theory, provides consistent and locally accurate feature attributions. We used a SHAP TreeExplainer on a random sample of 5000 test instances, calculating SHAP values for the positive class (‘local winner’). Our analysis of the SHAP outputs focused on three aspects:

• Global Feature Importance: The mean absolute SHAP value per feature, which ranks overall predictive impact.
• Feature Effects: The distribution of SHAP values, showing how feature values affect prediction magnitude and direction.
• Dependence and Interactions: The marginal effect of a feature on prediction, used to reveal non-linearities and potential interactions.

These quantitative analyses provide the basis for interpreting the model’s behavior.

3.5. Fuzzy Linguistic Interpretation of SHAP Values

To enhance the interpretability of the quantitative SHAP values, we introduce a fuzzy linguistic interpretation layer. This involves defining linguistic variables and their corresponding fuzzy sets based on a data-driven heuristic analysis of the observed distribution of SHAP values from our model. This approach ensures that the linguistic categories are directly relevant to the scale of impacts produced by the model. A detailed justification for the threshold selection and a discussion of its robustness are provided in Appendix A, and details of fuzzy parameter selection for linguistic interpretation can be found in Appendix B.

The impact of an individual feature’s SHAP value (denoted $s_j$) on the prediction is categorized using a linguistic variable “Influence Strength”, with five fuzzy sets: “Very Low”, “Low”, “Moderate”, “High”, and “Very High”. These are defined over the range of absolute SHAP values $|s_j|$ using triangular membership functions $\mu (x;a,b,c)=max(0,min({\displaystyle \frac{x-a}{b-a}},{\displaystyle \frac{c-x}{c-b}}))$ with the following empirically determined parameters:

• $P_1=0.0,P_2=0.1,P_3=0.5,P_4=1.0,P_5=2.0,P_6=3.0,P_7=4.5$.

The direction of influence is captured by the sign of the SHAP value $s_j$:

• Positive SHAP value ($s_j>ϵ$, for a small threshold $ϵ$): Indicates “Positive Influence” (increased likelihood of a local winner).
• Negative SHAP value ($s_j<-ϵ$): Indicates “Negative Influence” (decreased likelihood of a local winner).
• Near-zero SHAP value ($|s_j| \le ϵ$): Indicates “Neutral Influence”.

The global feature importance, calculated as the mean absolute SHAP value (${\overline{s}}_k$), is categorized using the linguistic variable “Overall Impact”. Its terms (“Negligible”, “Low”, “Medium”, “High”, “Very High”) are defined with the following empirically determined parameters:

• $Q_1=0.0,Q_2=0.02,Q_3=0.07,Q_4=0.13,Q_5=0.20,Q_6=0.35,Q_7=0.60$.

For individual predictions, the linguistic term for magnitude with the highest membership value is combined with its direction (e.g., “Moderate Positive Influence”). For global importance, features are categorized by their “Overall Impact” level. This provides a qualitative summary that is easier to communicate, with the choice of triangular membership functions being common due to their simplicity and interpretability [48].

4. Results

This section details empirical findings: model performance and SHAP insights into outcome drivers.

4.1. Model Performance Comparison

To evaluate the effectiveness of our chosen modeling approach, we benchmarked the LightGBM model against two standard baseline classifiers: a logistic regression and a random forest. All models were trained and tested on the identical feature set and data splits to ensure a fair comparison. The performance metrics, summarized in

Table 2. Comparative performance of predictive models on the test set.

Model	AUC	Accuracy	F1-Score
Logistic Regression	0.717	0.653	0.772
Random Forest	0.810	0.702	0.808
LightGBM (Proposed)	0.855	0.913	0.953

Note: Metrics for logistic regression and random forest are calculated at the default 0.5 classification threshold. The LightGBM metrics are reported at its optimized threshold (0.1387) to represent its peak performance, as detailed in Section 4.2.

, demonstrate that the LightGBM model achieves superior predictive power across all key metrics. This superior performance is a crucial prerequisite for building a reliable explanatory model, as a model that better captures the underlying patterns in the data can provide more trustworthy insights.

4.2. Predictive Model Performance

The LightGBM model effectively distinguished local from non-local awards, achieving an AUC of 0.8552 on the test set, indicating good class separability.

Classification performance was examined at the default 0.5 threshold and an optimal threshold. The optimal threshold to maximize accuracy was found to be 0.1387 (achieving an accuracy of 0.9129), while the optimal threshold to maximize the F1-score for the positive class was 0.1288 (achieving an F1-score of 0.9534). Figure 1 illustrates the performance of different metrics across various thresholds. For subsequent reporting of optimal performance, we use the threshold of 0.1387.

Figure 2 shows the confusion matrices. The default threshold (Figure 2a) results in an accuracy of 0.7520. The optimal threshold of 0.1387 (Figure 2b) significantly improves overall accuracy to 0.9129 and the F1-score for the local class to 0.9534 by adjusting the decision boundary, which is particularly relevant given the class imbalance.

Table 3. Classification Report Summary on Test Set.

Threshold	Class	Prec.	Rec.	F1	Support
Default (0.50)	Non-Local (0)	0.253	0.793	0.383	124,516
	Local (1)	0.971	0.748	0.845	1,157,782
	Accuracy			0.752	1,282,298
Optimal (0.1387)	Non-Local (0)	0.647	0.226	0.335	124,516
	Local (1)	0.922	0.987	0.953	1,157,782
	Accuracy			0.913	1,282,298

details precision, recall, and F1-scores. The optimal threshold (0.1387) substantially improves recall for the minority non-local class (Class 0) compared to what it would be at default, while maintaining high performance for the majority local class (Class 1). The choice of threshold depends on specific application priorities (e.g., prioritizing detection of non-local winners vs. overall accuracy).

4.3. Explainability Insights (SHAP Results)

SHAP values were calculated from the LightGBM model using a sample of 5000 test instances. These quantitative insights are further clarified using the fuzzy linguistic framework defined in Section 3.5.

Global Feature Importance with Fuzzy Linguistic Labels: Figure 3 shows the mean absolute SHAP value for each feature, indicating overall predictive power. Applying our fuzzy linguistic variable “Overall Impact”, historical_local_win_rate (mean abs SHAP ≈ 0.53) and AWARD_VALUE_EURO_FIN_1_log (mean abs SHAP ≈ 0.31) demonstrate a “Very High Overall Impact”. Features like ISO_COUNTRY_CODE (≈0.18), MAIN_ACTIVITY (≈0.17), and CAE_TYPE (≈0.14) exhibit a “High Overall Impact”. Others such as no_competition (≈0.12), NUMBER_OFFERS_log (≈0.11), and TYPE_OF_CONTRACT (≈0.10) show “Medium Overall Impact”. The remaining features display “Low” to “Negligible Overall Impact”. A summary is provided in

Table 4. Fuzzy linguistic interpretation of top global feature importance (values are approximate from Figure 3).

Feature Name	Mean Abs SHAP	Fuzzy Overall Impact
historical_local_win_rate	≈0.53	Very High
AWARD_VALUE_EURO_FIN_1_log	≈0.31	Very High
ISO_COUNTRY_CODE	≈0.18	High
MAIN_ACTIVITY	≈0.17	High
CAE_TYPE	≈0.14	High
no_competition	≈0.12	Medium
NUMBER_OFFERS_log	≈0.11	Medium
TYPE_OF_CONTRACT	≈0.10	Medium
country_tender_count	≈0.09	Medium
AWARD_VALUE_EURO_FIN_1_sqrt	≈0.08	Low

.

Feature Effects and Interactions (Beeswarm and Dependence Plots): The SHAP summary plot (Figure 4) details SHAP value distributions. The SHAP dependence plots in Figure 5 provide more granular insights into feature effects, which are interpreted here using our fuzzy linguistic framework:

• historical_local_win_rate (Figure 5a): This top feature exhibits a very strong, almost linear positive relationship. Low historical local win rates (e.g., below 0.7) exert a “Strong to Very High Negative Influence” (SHAP values often <−2.0) on predicting a current local win. Conversely, high historical rates (e.g., above 0.9) show a “Strong to Very High Positive Influence” (SHAP values often >+2.0). This suggests powerful path dependency. The interaction with no_competition (color) shows that for any given historical rate, the presence of no competition (red dots) tends to slightly increase the SHAP value (more positive).
• AWARD_VALUE_EURO_FIN_1_log (Figure 5b): As log-transformed award value increases, especially beyond a log value of ≈7–8, there is an increasingly “Strong to Very High Negative Influence” on predicting a local winner (SHAP values become more negative). For very high values (log > 12–15), this negative influence is substantial. Lower values show more dispersed, “Neutral to Slight Influence”. The interaction with ISO_COUNTRY_CODE (color) indicates that the general trend holds across countries, but the exact magnitude of SHAP values can vary.
• ISO_COUNTRY_CODE (Figure 5c): This plot shows the SHAP values for different (label-encoded) buyer countries. There is considerable variation. Some countries (e.g., with codes around 25–30 in the plot) tend to have negative SHAP values, indicating a “Slight to Moderate Negative Influence” on local awards (i.e., more openness to non-local). Other countries show positive SHAP values, suggesting a “Slight to Moderate Positive Influence” towards local awards. The interaction with no_competition (color) shows that single-bid situations (red dots) generally push SHAP values higher (more positive/less negative) for most countries.
• CAE_TYPE (Figure 5d): Contracting Authority Type shows varied influence. For instance, CAE_TYPE ‘5’ (European institutions) consistently shows “Very High Negative Influence” (large negative SHAP values). Other types like ’1’ (National) or ‘3’ (Regional) often have “Slight to Moderate Positive Influence”. CAE_TYPE ‘4’ (Public Law Bodies) shows high variance. Interaction with NUMBER_OFFERS_log (color) indicates that higher competition (blue dots) tends to lower SHAP values (more negative/less positive) across most CAE types.
• NUMBER_OFFERS_log: Based on the summary plot (Figure 4), very low numbers of offers (log value 0, i.e., single offer, related to no_competition) tend to have a “Moderate Positive Influence”. As the number of offers increases (log value > 1–1.5), the influence generally becomes negative (“Slight to Moderate Negative Influence”), suggesting that more competition favors non-local winners.
• no_competition (Figure 5e): The presence of no competition (value 1) typically results in a “Moderate Positive Influence” on predicting a local winner (SHAP values mostly between 0 and +1.0). When there is competition (value 0), SHAP values are generally negative. Interaction with ISO_COUNTRY_CODE shows that this effect is fairly consistent across countries.

5. Discussion

This section discusses findings regarding market integration, potential biases, smart governance, limitations, and future research.

5.1. Interpreting Model Findings in the Context of Market Integration

RQ1 sought the key drivers of local vs. non-local awards. SHAP identified historical_local_win_rate and log-contract value (AWARD_VALUE_EURO_FIN_1_log) as paramount (Figure 3), both exhibiting a “Very High Overall Impact”. Higher contract values strongly decrease local award probability (Figure 5b), translating to a “Strong Negative Influence” on local awards, especially for very large contracts. This aligns with economic intuition: scale fosters integration by attracting international bidders and justifying the transaction costs of cross-border participation [55]. The historical_local_win_rate demonstrates a powerful path-dependency effect (Figure 5a), where past national tendencies strongly predict current outcomes.

Regarding competition, the data reveals nuanced patterns. The binary feature no_competition (single bids) shows a clear “Moderate Positive Influence” favoring local winners (Figure 5e). As competition increases (reflected partly by NUMBER_OFFERS_log, which has a “Medium Overall Impact”), the tendency generally shifts towards favoring non-local winners, as suggested by the SHAP summary plot (Figure 4).

Addressing RQ2, significant heterogeneity exists. The substantial importance of features like ISO_COUNTRY_CODE (Figure 5c), MAIN_ACTIVITY, and CAE_TYPE (Figure 5d) indicates uneven market openness. SHAP dependence plots reveal that these categorical features have distinct influence patterns. For instance, certain contracting authority types (e.g., ‘5’—European institutions) show a “Very High Negative Influence” toward local awards, while others (e.g., ‘1’—National authorities) often exert a “Slight to Moderate Positive Influence” favoring local winners. Similarly, different countries (ISO_COUNTRY_CODE) exhibit varying baseline tendencies.

The fuzzy linguistic layer aids in communicating these strengths of influence. For example, stating that contract value has a “Very High Overall Impact” and typically exerts a “Strong Negative Influence” on local awards for large contracts is more immediately understandable for policymakers than citing raw SHAP values alone.

5.2. Exploring Potential Biases and Fairness Implications

Observed heterogeneity, especially from features like CAE_TYPE or B_GPA after controlling for economic factors, raises questions about systemic biases (RQ2). The fuzzy interpretation helps to qualitatively grade the strength of these tendencies. If a buyer type consistently shows a “Moderate or High Positive Influence” towards local winners, it might suggest practices disadvantaging non-local bidders, conflicting with Single Market principles. Caution is crucial: SHAP reveals model prediction drivers based on correlations, not causality. A local award tendency might stem from legitimate unmodeled factors (e.g., specialized local needs) not overt bias. The model could also perpetuate historical data biases [39]. SHAP, with its fuzzy linguistic enhancement, offers diagnostics for fairness investigations but does not definitively prove bias without further analysis.

5.3. Implications for Smart Governance and Policy

This interpretable framework has significant practical implications for smart governance and evidence-based policy (RQ3).

First, it offers a powerful monitoring tool for procurement oversight. Authorities can compare model predictions and SHAP values for new tenders against actual outcomes to flag anomalous awards for scrutiny. Particularly, tenders with high contract values (AWARD_VALUE_EURO_FIN_1_log) and numerous offers (NUMBER_OFFERS_log) that nevertheless go to local winners represent statistical outliers that might warrant closer examination, as these characteristics typically exert a “Strong Negative Influence” on local award probability.

Second, global feature importance and dependence plots offer actionable insights for policy interventions. Identifying specific buyer types (CAE_TYPE) or sectors (MAIN_ACTIVITY) with consistently low international openness allows for targeted capacity-building, training programs, or enhanced transparency requirements. For instance, contracting authority types showing a “Moderate to High Positive Influence” toward local winners might benefit from specialized support for managing international procurement processes. Understanding the impact of award criteria types (CRIT_CODE) can prompt reviews of whether certain evaluation approaches inadvertently disadvantage non-local bidders.

Third, the model quantifies the impact of existing regulations and procurement framework characteristics. For example, the influence of government procurement agreement coverage (B_GPA) can be described with terms like “Moderate Negative Influence” on local awards in certain contexts, providing accessible feedback on the effectiveness of international procurement agreements. This helps policymakers assess whether such frameworks are achieving their intended goals of market openness.

Fourth, the findings suggest concrete governance applications:

• Targeted oversight: Monitoring efforts should focus on high-value contracts with many offers that unexpectedly go to local winners, as these represent statistical anomalies based on our model.
• Benchmarking contracting authorities: The significant CAE_TYPE effect suggests creating a data-driven ranking of authority types by their tendency toward local awards, controlling for tender characteristics. This could help identify best practices from the most internationally open buyer categories.
• Procedural reforms: The substantial influence of categorical features like award criteria (CRIT_CODE) suggests reviewing how qualification requirements and evaluation approaches might inadvertently advantage local firms.

Fifth, the fuzzy linguistic interpretation layer makes these insights more readily communicable to diverse stakeholders, fostering a shared understanding essential for collaborative governance. By translating complex SHAP values into intuitive terms like “Strong Positive Influence” or “Moderate Negative Influence”, the findings become accessible to non-technical policymakers, procurement professionals, and oversight bodies. This common interpretative framework supports more inclusive policy discussions around market integration and procurement fairness.

In summary, this XAI approach transforms raw procurement data into active intelligence for adaptive oversight and policy improvement. It moves beyond simple descriptive statistics to offer deeper, more nuanced insights into the complex factors driving cross-border procurement outcomes, while ensuring that these insights remain understandable to all stakeholders through the fuzzy linguistic framework.

5.4. Limitations

This study has limitations. Data: Analysis relies on awarded CANs in TED, excluding unsuccessful/cancelled tenders, potentially biasing the sample. TED data has inconsistencies and missing values; imputation handles missingness statistically. Lack of detailed bidder information (firm size, capabilities) limits control for firm-level factors. Model: LightGBM learns from data, including potential biases. Predictions are correlational, not causal. SHAP, while advanced, has assumptions and explains model behavior, an approximation of reality [29]. Scope: Analysis is for 2018–2023 and treats EU market homogenously (though ISO_COUNTRY_CODE captures some country effects). Deeper regional analysis or focus beyond winner nationality (e.g., SMEs) could offer more insights.

Fuzzy threshold selection: The fuzzy linguistic framework relies on thresholds selected via a data-driven heuristic approach, as detailed in Appendix A. While this method ensures that the thresholds are relevant to our model’s output, their exact placement is a methodological choice. The qualitative conclusions of this study, however, are robust to minor variations in these breakpoints, as they are based on the large, underlying differences in SHAP values between high- and low-impact features. Future research could explore more formalized methods for threshold selection, such as statistical clustering or expert elicitation, to further refine the linguistic mappings.

5.5. Methodological Contribution

The comparative experiments (Section 4.2) serve to validate the choice of LightGBM as a robust foundation for our explanatory framework. Focusing on the threshold-independent AUC metric, which provides the fairest comparison of inherent predictive power, our model (0.855) demonstrates a significant advantage in discriminative capability over both logistic regression (0.717) and random forest (0.810). This superior predictive performance is a critical prerequisite; a model that cannot reliably capture the underlying patterns in the data cannot be trusted to provide meaningful explanations.

With the model’s high performance established, our primary methodological contributions unfold. Our approach moves beyond simple correlation to the attribution of effects, using SHAP to quantify the “local advantage” or disadvantage associated with specific procurement practices. We also address the practical challenge of class imbalance typical in procurement data through the optimization of the classification threshold, which allows for achieving high accuracy (0.9129) while understanding trade-offs in class detection (Table 3). Perhaps most significantly, we introduce a novel fuzzy linguistic interpretation layer over the SHAP analysis. By translating numerical SHAP values into qualitative descriptors like “Strong Positive Influence”, we bridge the gap between technical precision and human understanding, a combination particularly valuable in governance applications.

This methodological approach offers a template for analyzing other complex policy domains where administrative data contains mixed features, class imbalance challenges, and the need for both technical precision and interpretable outputs for non-technical audiences.

5.6. Future Research Directions

Future research could enrich datasets with bidder-specific information (e.g., from Orbis) and finer geographical data (NUTS levels). NLP on tender text could extract nuanced features from technical specifications and award criteria descriptions that might influence cross-border participation. Longitudinal analysis could track evolving patterns over time, especially in response to policy changes.

Several promising research extensions emerge from our findings:

• Counterfactual analysis: The model could be used to simulate “what if” scenarios—for example, estimating cross-border procurement rates if all tenders followed the practices of the most internationally open buyer types or if certain policy changes were implemented.
• Pre/post policy evaluation: This XAI framework could evaluate the effectiveness of EU procurement directives and initiatives by analyzing changes in feature importance and SHAP values over time, particularly before and after major policy implementations.
• Tender text analysis: Combining our approach with natural language processing could reveal whether specific language patterns or terminology in tender documents correlates with local awards, potentially identifying subtle linguistic barriers to cross-border participation.
• Network effects: Extending the analysis to include network relationships between contracting authorities and suppliers could reveal patterns of repeated interactions and their influence on cross-border procurement.

Methodologically, future work should move toward causal inference (e.g., propensity score matching, regression discontinuity designs) to isolate the impacts of specific policies or procurement characteristics. This would provide stronger evidence for policy interventions by distinguishing causation from correlation. Additionally, refining the fuzzy linguistic interpretation framework through stakeholder feedback could further enhance the usability of XAI outputs for diverse users in the procurement ecosystem.

6. Conclusions

This study aimed to identify factors influencing local versus non-local public contract awards in the EU, an indicator of Single Market integration, using interpretable machine learning (LightGBM with SHAP) on 2018–2023 TED data.

The model achieved robust predictive performance (AUC = 0.8552; accuracy = 0.9129 and F1-score for local class = 0.9534 at an optimal threshold of 0.1387). SHAP analysis revealed that historical_local_win_rate and higher contract values (AWARD_VALUE_EURO_FIN_1_log) are the most influential predictors, both categorized linguistically as having a “Very High Overall Impact”. Larger contracts significantly increase non-local winner likelihood, typically exerting a “Strong Negative Influence” on local awards.

Competition levels also play a key role, with single-bid procedures (no_competition) showing a “Moderate Positive Influence” toward local winners, while increased competition generally favors non-local suppliers. Structural factors like buyer country (ISO_COUNTRY_CODE), contracting authority type (CAE_TYPE), and main activity (MAIN_ACTIVITY) exert substantial influence (“High Overall Impact”) despite controlling for economic variables, indicating systemic differences in market openness.

The key contribution is demonstrating XAI’s value in public procurement governance, further enhanced by a fuzzy linguistic interpretation layer that translates complex quantitative outputs into accessible, human-understandable terms. This combined approach provides transparent, interpretable evidence to inform policy, assess market integration nuance, identify potential barriers or fairness concerns, and support data-driven monitoring.

As the EU pursues a more integrated, fair, and efficient Single Market, this interpretable AI approach, through its fuzzy linguistic framework, offers a potent path to enhanced transparency and accountability in public spending. It modernizes procurement oversight by moving beyond simple descriptive statistics to deeper, more nuanced understanding, while ensuring that these insights remain accessible to all stakeholders.