Integrating Structural Causal Model Ontologies with LIME for Fair Machine Learning Explanations in Educational Admissions

: This study employed knowledge discovery in databases (KDD) to extract and discover knowledge from the Benue State Polytechnic (Benpoly) admission database and used a structural causal model (SCM) ontological framework to represent the admission process in the Nigerian polytechnic education system. The SCM ontology identified important causal relations in features needed to model the admission process and was validated using the conditional independence test (CIT) criteria. The SCM ontology was further employed to identify and constrain input features causing bias in the local interpretable model-agnostic explanations (LIME) framework applied to machine learning (ML) black-box predictions. The ablation process produced more stable LIME explanations devoid of fairness bias compared to LIME without ablation, with higher prediction accuracy (91% vs. 89%) and F1 scores (95% vs. 94%). The study also compared the performance of different ML models, including Gaussian Naïve Bayes, Decision Trees, and Logistic Regression, before and after ablation. The limitation is that the SCM ontology is qualitative and context-specific, so the fair-LIME framework can only be extrapolated to similar contexts. Future work could compare other explanation frameworks like Shapley on the same dataset. Overall, this study demonstrates a novel approach to enforcing fairness in ML explanations by integrating qualitative SCM ontologies with quantitative ML/LIME methods.

This study focuses on using KDD to extract knowledge from Benue State Polytechnic's (Benpoly) admission database in Nigeria.Previously, the admission process at Benpoly relied on the admissions committee's manual evaluation of applicant data.This process was timeconsuming, prone to human biases, and lacked transparency in decision-making.Furthermore, there was no standardized method for representing the admission criteria and their interrelationships.Previous studies have developed ontologies for university admission processes using tools like Protégé [50]- [54], but the context, methods, and focus differ from this study, which develops an application-based structural causal model (SCM) ontology for the Nigerian polytechnic admission system.SCM ontologies have been used in other studies to simulate randomized control trials and analyze causal impacts [55]- [58], but not for the purpose of this study.
Processing educational databases for admission purposes presents several general challenges: (i) Data quality: Educational databases often contain missing values, inconsistencies, and errors that need to be cleaned and preprocessed before analysis, (ii) Feature selection: Identifying the relevant features that impact admission decisions from a large number of variables can be challenging and requires domain expertise, (iii) Data integration: Admission databases may need to be integrated with data from other sources, such as entrance exam scores or demographic information, which can be complex due to differences in data formats and structures, (iv) Privacy and security: Working with educational data involves handling sensitive personal information, which requires strict adherence to data privacy regulations and security measures.Our work on the Benpoly admission database addresses these challenges through a comprehensive data preprocessing pipeline, as detailed in the "Data Preprocessing" of section 3.2.This includes data cleaning, feature engineering, outlier removal, and data transformation steps specific to the Benpoly database structure and admission process.
To address this problem, this study proposes a novel fair-LIME framework that integrates a qualitative SCM ontology with the quantitative ML/LIME methods.The SCM ontology is used to identify biased features in the dataset and constrain or ablate them [59], [60] before inputting them into the ML/LIME process to generate faithful and fair explanations.
Thus, the main contributions of this study are: (1) Applying KDD to mine knowledge from a polytechnic admission database, (2) Developing an SCM ontology to model the causal relations in the admission process, (3) Validating the SCM ontology assumptions using the CIT, (4) Using the SCM to identify biased features and constrain the LIME framework, (5) Developing a novel "fair-LIME" explanation approach that balances fidelity and fairness.
To the best of our knowledge, this is the first study to integrate a qualitative SCM ontology with quantitative machine learning and LIME techniques to enable fairer explanations of a real-world admission process.The proposed framework has the potential to be extended to similar admission contexts.
The rest of the paper is structured as follows: Section 2 reviews related work, Section 3 describes the materials and methods used, Section 4 presents the implementation and results, and Section 5 concludes the paper and suggests future research directions.

Related Works
Machine learning (ML) frameworks are used to build predictive models that learn patterns and relationships from data.These models can be complex and opaque, making understanding how they arrive at their predictions difficult.On the other hand, LIME (Local Interpretable Model-agnostic Explanations) is an explanation framework that provides posthoc interpretability for black-box ML models.LIME generates local explanations for individual predictions by approximating the behaviour of the complex ML model with a simpler, interpretable model (e.g., linear regression) in the vicinity of the instance being explained.These explanations highlight the most influential features of a particular prediction [61].
In this study, ML and LIME frameworks complement each other in the following ways: (1) ML models make predictions, while LIME provides explanations, (2) LIME uncovers biases in ML models, (3) SCM ontology and ablation improve LIME explanations, and (4) LIME evaluates the impact of ablation.Gaussian Naïve Bayes (GNB) was chosen due to its high prediction accuracy when implemented with the LIME explanation framework in this study when compared with other ML algorithms [62]- [66].
Even though no known ontologies specifically represent the admission process in the Nigerian polytechnic education system, some ontologies have been developed for university admission processes using tools like Protégé [50]- [54].However, these ontologies differ in context, method, and focus from the current study, which employs a SCM technique to design an application-based ontology for the polytechnic admission process.
Recent studies have designed and validated application-based SCM ontologies using the conditional independence test (CIT) criteria for simulating randomized control trials and analyzing causal impacts in primary schools [55]- [58].In contrast, this study uses the SCM ontology to identify causal relations in features of a polytechnic admission dataset, which are then used to constrain and sparse the quantitative ML and LIME frameworks to enable fair explanations.
LIME is a model-agnostic technique that provides local explanations by approximating the behaviour of a complex model with an interpretable surrogate model around a specific instance [61].However, LIME explanations can be biased when based on patterns learned from datasets that violate fairness [67]- [73].Ablation, the process of systematically removing or deactivating parts of a model or system, has been used to understand the contributions or effects of different components on overall performance or behaviour [59], [60], [74]- [76].
This study addresses the research gap of integrating qualitative SCM ontologies with quantitative ML and LIME methods to enforce fairness in explanations.By using the SCM ontology to identify biased features and constrain them through ablation before inputting them into the ML/LIME process, the proposed fair-LIME framework generates explanations that are faithful to the underlying causal relations and devoid of fairness bias.
While previous studies have developed ontologies for university admission processes and used SCM ontologies for causal impact analysis, this study is novel in its application of SCM ontology to constrain ML/LIME frameworks for fair explanations in the context of the Nigerian polytechnic admission process.

Material and Methods
This section describes all the materials and methods that have been implemented in this study.

Database Structure and Description
The dataset employed in this study was mined or extracted from Benpoly's admission web portal database.Table 1 summarizes the important variables and features used in the study, including their abbreviations, descriptions, data types, and roles in the admission process.Throughout this work, we refer to the variables and features using their abbreviations (as in Table 1) to improve readability and conciseness.The full names are used only when necessary for context or emphasis.The database contained 23 variables set, which were both alphanumeric and contained 12,043 records.Figure 1 shows an image excerpt from the admission database, while Figure 2 shows a comprehensive ontological framework of the entire admission database, with classes, attributes, and their relations.

Data Preprocessing
Figure 3 shows the data preprocessing stages that were applied to the admission database and further used for the KDD process of ML prediction and LIME explanations.The raw Benpoly admission dataset contained 12,043 records and 23 variables.The preprocessing steps and their impact on the dataset are as follows: 1.Data cleaning and feature selection: 8 variables that were not relevant for modeling the admission process were removed.Records with missing values in the remaining variables were also dropped.After this step, the dataset contained 11,869 records and 15 variables.2. Feature engineering: A new variable called "Current_Qualification" was created based on the existing "Course_Category" variable better to represent the applicants' eligibility for different admission types.This increased the number of variables to 16.The original "Course_Category" variable was then dropped, keeping the dataset at 15 variables.3. Outlier removal: 174 records with outlier values in the "Age" variable (applicants with ages below 15 or above 70 years) were identified and removed.After this step, the dataset contained 11,695 records and 15 variables.4. Data transformation: The categorical variables were converted to numerical format using one-hot encoding.This increased the number of variables to 151, while the number of records remained at 11,695.
Table 2 provides an overview of the dataset size and composition changes throughout the preprocessing pipeline.In summary, the data preprocessing steps resulted in a final dataset with 11,695 records and 151 variables used for subsequent analysis and modeling.Figure 3 visualizes the KDD data analytics lifecycle followed in this study.

Dataset Division for Model Training and Testing
After the data preprocessing steps, the final dataset contained 11,695 records and 151 variables.This dataset was divided into training and testing sets using a stratified random sampling approach to ensure that the class distribution (admitted vs. not admitted) was preserved in both sets.Specifically, the train_test_split function from the scikit-learn library in Python was used with the following parameters: 1. test_size = 0.3: This allocates 30% of the records (5,146 records) to the testing set and the remaining 70% (8,186 records) to the training set. 2.
random state = 0: This sets a fixed random seed for reproducibility, ensuring that the same split is obtained in repeated runs of the code.3. stratify = y: This ensures that the class distribution in the original dataset is maintained in both the training and testing sets.Here, 'y' refers to the target variable (Admission_Status).
The resulting training set contained 8,186 records (70%), while the testing set contained 5,146 records (30%).The training set was used to train the Gaussian Naive Bayes (GNB) classifier, and the alternative models (Decision Trees and Logistic Regression) were used for comparison.The trained models were then evaluated on the testing set to assess their performance and generalizability to unseen data.The 70%-30% split was chosen as it is a commonly used ratio in machine learning that provides a good balance between having enough data for training and enough data for testing the model's performance on unseen examples.The stratified sampling approach ensures that the model is trained and evaluated on a representative distribution of the target classes, mitigating potential biases arising from imbalanced splits.
Furthermore, the same training and testing sets were used for all the models and experiments in this study to ensure a fair and consistent comparison of their performance and the impact of the Fair-LIME framework.

Structural Causal Model Ontological Framework
The SCM ontological framework is a formalism used in causal inference and modeling to represent causal relationships among variables in a system.It provides a mathematical and graphical framework for expressing causal hypotheses and identifies important variables used mostly in making causal inferences based on observational or experimental data [77]- [80].In SCM, variables represent the entities or factors of interest in the studied system.These variables can be observed or unobserved and can take on different values.Further, in SCM, structural equations are used to describe the functional relationships between variables in the system.Each structural equation specifies how the value of a variable depends on the values of its parent variables in the causal graph or the direct acyclic graph (DAG) as it is called.For example, if a variable  is a child variable depending on parent variables  1 ,  2 , …   the structural equation for Y will be written in Equation (1).
Where   =  1 ,  2 , …   and the function (. ) represents a deterministic or stochastic function, and ∈  represents an error term capturing the unmodeled influences or noise within the dataset.
The DAG represents the causal relationships between variables in the system.It consists of nodes representing variables and directed edges or arrows representing causal dependencies between variables.The absence of a direct edge between two variables indicates that they are conditionally independent given their parents in the graph, and that is the basis for the statistical CIT criteria used in the study.Other attributes of the SCM that do not apply to this study are Interventions and Counterfactuals [81], [82].Thus, the SCM framework provides a principled approach to causal reasoning and inference, allowing researchers to formalize causal hypotheses, make predictions about the effects of interventions, and test causal hypotheses using observational or experimental data.By explicitly representing causal relationships in a DAG, the SCM framework helps identify important variables to focus on in a data distribution in a particular task, thus enabling more robust and reliable causal conclusions or other related analysis.Therefore, this work employs the SCM ontological framework to identify relevant variables in the admission database that are key in determining the admission process in the polytechnic education system in Nigeria and further uses the identified features for the KDD modeling, prediction, and LIME explanations.

LIME Explanation Framework
The LIME framework is a model-agnostic technique employed in explaining a data instance in a black-box ML prediction model.LIME offers a local explanation by approximating the behavior of a complex model with an interpretable surrogate model, such as a linear model, around a specific instance of interest [61].This involves: Selecting an Instance for which you desire to explain the prediction made by the model.This instance could be a single data point or observation from the dataset.Next, LIME generates perturbations or variations of the instance by randomly sampling from the neighborhood of the instance.These perturbations introduce slight changes to the features while keeping the target instance close in terms of similarity.
Further, for each perturbed instance, LIME obtains predictions from the complex blackbox model that we aim to explain.These predictions serve as the ground truth to which the surrogate model will be fitted.Then, LIME fits an interpretable surrogate model (e.g., linear regression) to the perturbed instances and their corresponding predictions obtained from the complex model.The surrogate model aims to approximate the complex model's behavior in the target instance's local neighborhood.Once the surrogate model is trained, LIME interprets the coefficients or weights assigned to each feature in the model.These coefficients indicate each feature's relative importance or contribution to the prediction made by the complex model around the target instance.Finally, LIME generates explanations for the prediction of the complex model by highlighting the features that have the most significant influence on the prediction according to the surrogate model.These explanations help users understand which features drive the model's decision for the selected instance.Equation ( 2) below shows the LIME explanation framework.
Where () is the data instance explanation,  is the complex predictive model, and  is a simple interpretable surrogate model, and  ∈ , where  is a class of sparse interpretable models, such as linear models, decision trees, falling rule lists, etc., [83].
The first loss term ℒ(, ,   )zin the optimization function means we look for the approximation of the complex prediction model  by the simple and sparse model , in the neighborhood of the focused dataset point   (which is a proximity measure).The second loss term Ω() is used to regularize the complexity of the simple surrogate interpretable model  (e.g., reducing the depth of a tree in a decision tree or the number of non-zero weights for a linear regression model to enable sparseness and comprehension for people).The Lasso Regression regularization technique is used in practice to implement the Ω() term [65].Thus, ensuring a simple explanation with only a few relevant variables.
Hence, the loss term ℒ(, ,   ) is calculated using Equation ( 3) by a method called perturbation.
Where the () is the label or prediction target of the complex prediction model, and ( ′ ) is the predictions from the simple interpretable surrogate model  (which comes from the perturbed features), and the term   () weights the loss function of the perturbed features according to the proximity of the data point vis-à-vis the threshold set by the complex model prediction ().So, the perturbed features close to the original data point are weighted the most, and vice versa.Thus, enduring the local faithfulness or local fidelity of the model.

Fairness and Fidelity in Explanation Frameworks
Fidelity in machine learning explanations refers to the degree to which the explanations accurately represent the model's underlying behavior and decision-making process [67]- [69].Fidelity ensures that the explanations faithfully capture the logic, patterns, and relationships the model learns from the training data.Fidelity can be evaluated by comparing the explanations provided by the model to the actual model predictions or decisions [70].High-fidelity explanations should closely align with the model's outputs and reflect the same reasoning and decision-making process as the model.The LIME explanation framework is built on fidelity [61].
On the other hand, fairness in machine learning explanation frameworks relates to the accuracy, trustworthiness, and ethical considerations of the explanations provided by machine learning models.Fairness in machine learning explanations refers to the degree to which the model's predictions or decisions are free from biases, discrimination, and unfairness against certain individuals or groups [71]- [73], [84].Fairness considerations are crucial to ensure that machine learning systems do not perpetuate or amplify existing societal biases and disparities [85].Thus, Fairness in explanations involves providing insights into how the model treats different subgroups within the dataset and whether there are disparities or biases in the explanations provided for different groups.Explanations should highlight any potential biases or unfairness in the model's predictions or decisions and enable stakeholders to address these issues effectively.Fairness can be evaluated using various qualitative metrics and criteria, such as demographic parity, equal opportunity, and disparate impact analysis [86]- [88].These metrics assess whether the model's predictions or decisions exhibit fairness across different demographic groups, such as race, gender, age, or socioeconomic status.Thus, in this study, we seek to balance these two concepts of fidelity and fairness in LIME explanations in the mined Benpoly admission dataset.

Ablation with ML/LIME Frameworks
In machine learning and model interpretation, ablation refers to systematically removing or deactivating parts of a model or system to understand their contributions or effects on the overall performance or behavior [59], [60].Ablation studies are commonly used to analyze the importance of different components, features, or layers of a model and to assess their impact on model predictions or outputs [66], [89].The term "ablation" originates from medical science, where it refers to the surgical removal or destruction of tissue, organs, or body parts, often for therapeutic or experimental purposes [74], [90].In machine learning, ablation studies serve a similar purpose by "surgically" dissecting the model to understand its internal workings and dependencies.Thus, ablation studies are valuable for understanding machine learning models' robustness, interpretability, and possible generalization capabilities.They help identify critical components, features, or dependencies within the model architecture and can inform model improvement, feature selection, and interpretability efforts [75], [76].Hence, in this study, we will use the background knowledge of the dataset to design an applicationbased SCM ontological framework that can enable us to determine or identify the features relations in the dataset and that can be ablated and constrain the ML/LIME framework to enable us to generate ML/LIME predictions and explanations that are faithful and fair.

Our Fair-LIME Explanation Framework
In our proposed fair-LIME framework, the complex model predictive model  in equations 2 and 3 of the LIME explanation framework is constrained using the ablation technique.Hence making the complex model sparse.This is achieved by using the background knowledge from which the SCM ontology for the dataset is designed and then identifying the input features that are bound to cause fairness bias in the ML prediction and the concomitant LIME explanations.Thus, removing or constraining them (the ablation process) would become imperative to obtain predictions and explanations that are both faithful (meet fidelity) and fair.Thus, the constrained and sparse model   is then used as input into the LIME framework to replace the initial complex prediction model of Equations ( 2) and (3), as shown in Equations ( 4) and (5).
Where   is the constrained and sparse predicted model.Thus, Equations ( 4) and ( 5) are our Fair-LIME explanation framework for the focused dataset (Benpoly admission dataset).

Implementation and Results
This section uses the material (the Benpoly admission dataset) and the methods (SCM, LIME, Ablation, etc.) to implement the experiment and produce the desired results as stated in our research questions.

SCM Ontology Design for Benpoly Admission Dataset
Figure 6 shows the Structural Model (SCM) Ontological Framework for the BenPoly Admission process.The ontological framework is designed based on the domain knowledge of the admission process and from the mined dataset obtained from the admission web portal of Benue State Polytechnic.The initial dataset contained 23 variables, both alphanumeric and 12,043 records.The preprocessing stages involve cleaning and dropping variables and removing records that are not relevant for modeling the admission process.Thus, the first preprocessing stage reduced the number of variables to 10 and the records to 11869.The second stage involved performing feature engineering on a variable to generate a new variable.Thus, with the aid of domain knowledge and feature engineering, the variables labeled Course_Category were able to generate another important variable called Current_Qualification, which is important for modeling the admission process for the institution that was not initially a part of the 23-variable set.After the Current_Qualification variable was created from the Course_Category variable, the Course_Category was dropped.Thus, the total number of variables required for modeling the admission process is 9. Out of these, five are categorized as students' characteristics and labeled in the ontological framework as X.These variables include: gender, marital_status, lga, and Age.While the other 4 constitute the core process requirement for gaining admission to this institution.These variables include Course_appplied, current_Qualification, Mode_of_entry, and Admission_status.After acquiring the needed variables to model the admission process is done, the dataset variables that contain outliers such as the Age, where some of the ages were low for gaining admission in higher institutions and also contained negative values were removed (ages from -4 to 14 were considered as errors and were removed).Further, the alpha variables were then converted to numeric.This process of converting the alpha variables is important for validating the structural model (SCM) ontological framework designed with the dataset, and only the numeric forms of the dataset are valid in the tools used for their validation process.Similarly, the numeration of the alpha variables is important for the final model prediction of the dataset.Overall, the KDD preprocess stages required for SCM ontological framework modeling and validation of the admission process in BenPoly institution are listed below: 1. Cleaning dataset -removing and dropping columns and records not required to model the admission process. 2.
Feature engineering -developing a feature that is required but not part of the initial dataset but is extremely essential for the modeling process of the admission process.3. Removal of outlier -removing unreasonable and unrealistic records, such as underaged records for the admission process.4. Numeration of the alpha variables -Converting alpha (categorical) variables to numeric variables.5. Design of SCM ontological framework to represent the dataset from the identified/engineered features in the dataset with the help of the background knowledge of the polytechnic admission system.Figure 4 the waterfall model conceptual framework for the implementation of the study, which process is itemized as follows: 4. Design the algorithm that explains the CIT validation process. 5. Use the SCM Ontological framework to identify biased features that would constrain or ablate the ML/LIME frameworks to enable fair and actionable LIME Explanations.

SCM Ontology Explanations of Variable Relations
The SCM ontological design is best for explicating the admission process in the polytechnic (Benpoly) because of its propensity to depict the causal relations within the dataset so long as the domain knowledge for the dataset is known; and also, its ability to validate the same with the dataset [55]- [58], [91].Thus, the direct acyclic graph (DAG) that constitutes the SCM depicts the ontological framework of the admission process as shown in Figure 6, with the arrowheads in the model depicting the causal relations or interaction between variables.Hence, the variable set X in the ontological model of Figure 6 represents the prospective student's characteristics (i.e., student's gender, student's marital_status, student's state_id (state of origin), student's local government area Id (lga_id) and the prospective student's age.Thus, X = {gender, martialstatus, stateId, lgaid, Age}.Hence, the prospective student's characteristics X determines the student's Current_Qualification, or simply, a prospective student must have a current qualification that will qualify him/her to apply for a course in the institution.Hence the causal arrow head points from X to the Current_Qualification.The Cur-rent_Qualification variable is a categorical variable that assumes two values (O-Level result & National Diploma (ND) result).Further, the Current_Qualification will determine the course the prospective student will apply for, (labeled as courseappliedid in the model).Thus, a prospective student with an O-level can only apply for a National Diploma (ND) course, while a prospective student with an ND can only apply for a Higher National Diploma (HND) course.The relationship between the Current_Qualification and courseappliedid is causally related in both directions, as shown in SCM in the ontological framework of Figure 6, which states that the course a prospective student applies for will determine his/her current qualification.Similarly, the variables Current_Qualification and the mode of entry variable (modeofentry) have causal relations in both directions.
The modeofentry is also a categorical variable with two values (i.e., JAMB and WITH-OUT JAMB).The acronym JAMB stands for Joint Admission and Matriculation Board, and it is a central examination body that conducts entry examinations for all prospective students wishing to gain fresh admission into any Nigerian higher institution [92].Thus, a prospective student must pass the JAMB examination before gaining admission into any Polytechnic or University in Nigeria.Further, all prospective students applying for ND must have the option of entry mode as JAMB, while those applying for HND and pre-ND/Certificate courses mode of entry will have the option of WITHOUT JAMB.Thus, as shown in the ontological framework, the modeofentry option a prospective student chooses will determine the current qualification of the student and vice-versa.Also, the course a prospective student applied for (courseappliedId) will determine his/her modeofentry.Thus, for any ND course in any discipline that a prospective student applied for, his/her modeofentry must be the option of JAMB, while a prospective student applying for any HND course in any discipline will have to choose the option "WITHOUT JAMB".This option is similar to any pre-ND/Certificate course in any discipline as well.Hence, the causal direction arrows only point from the couresappliedid to the modeofentry, not vice versa.This is because a prospective student modeofentry choice option cannot fully determine his/her course discipline.Finally, the modeofentry is the final criterion determining a prospective student's admission status.That is, whether or not a prospective student would be admitted.The admission status is also a categorical variable with two options -i.e."admitted" and "not admitted".Thus, if a student's JAMB score (for a candidate applying for ND) is within the school's acceptable scores (JAMB cutoff marks) or the ND result (for candidates applying for HND or pre-ND/Certificate courses), the prospective student's current qualification result aligns (O-level or ND results) with the courseappliedid, then the prospective student admission status becomes "admitted".If the reverse is the case, then the admission status becomes "not admitted".Figure 5 visualizes the data distribution of the 9 variables or features depicted in the SCM ontology after preprocessing, while Figure 6 evinces the relationships amongst the variables in the dataset with the SCM framework.
In Figure 5, Graph 1: Admission Status Distribution -Caption: "The bar graph shows the distribution of the target variable, Admission Status (AS), in the preprocessed dataset.The majority class is 'Admitted' (AS = 1), indicating that a higher proportion of applicants are granted admission."-Annotation: "Class imbalance: 70% Admitted (AS = 1), 30% Not Admitted (AS = 0)".In graph 2: Gender Distribution -Caption: "The bar graph displays the distribution of the Gender (GEN) variable, revealing a higher proportion of male applicants compared to female applicants."-Annotation:"Gender distribution: 60% Male (GEN = 1), 40% Female (GEN = 0)".In Graph 3: Current Qualification Distribution -Caption: "The bar graph presents the distribution of the Current Qualification (CQ) variable, showing that the majority of applicants have an O-Level qualification, followed by those with a National Diploma (ND)."-Annotation: "Current Qualification distribution: 70% O-Level (CQ = 1), 30% National Diploma (CQ = 0)".In   The SCM ontological framework is designed using the domain knowledge of the admission process and is implemented using three major tools.Which are: 1. Python Jupyther notebook: Used for the preprocessing and feature engineering.2. Digitty package: Used for the design of the ontological framework as seen in Figure 6 and for obtaining model coordinates and the CIT criteria (model implied assumptions) that exits among the variable set and uses the same in R programming for validating the model with the dataset.3. R programming: model coordinates and CIT criteria obtained from the Dagitty design package is imported into R and used alongside the dataset to validate the design model using the CIT criteria, a statistical technique for testing the condition independence that exists among the variables.
The code implementation of the entire process can be accessed at GitHub.

SCM Ontology Validation Results Presentation
The preprocessing and design processes are already explicated above with the final design of the SCM ontological framework from the Dagitty software package, as shown in Figure 6.Thus, this section presents the CIT criteria and results obtained from the validation process.
The validation process in R requires two major components: (i) the coordinates and CIT criteria obtained from the design process in Digitty and (ii) the dataset.Thus, the CIT criteria obtained from the structure of the SCM ontology is given in Equation 6or 6b below: As shown in the abridged outputted results of R in Table 1.Where X={gndr, mrtl, sttd, lgad, Age} Where the symbol "⊥" or "_||_" as shown in Table 1, stands for independent of, and "|" stands for, given or conditioned on.
Therefore, we can interpret Equation 6& 6b as follows: • Current_Qualifcation is independent of Admission_Status conditioned on modeofentry, and Courseappliedid is independent of Admission_Status conditioned on modeofentry, and

•
Courseappliedid is independent of X, not conditioned on any variable, and

•
Modeofentry is independent of X, not conditioned on any variable and.
• Admission_Status is independent of X, not conditioned on any variable.
Further, these identified CIT criteria will be used in R alongside the dataset to perform the CIT statistical test.This test aims to confirm or reject the CIT assumptions encoded and identified in the SCM ontological framework.If the CIT assumptions identified in the SCM and DAG are affirmed, then the SCM ontology structure is correct and validated.Otherwise, one will need to redesign it again or possibly check the dataset [93].
Thus, since the X in the model represents five variables, five different CIT are performed on equation 6 or 6b for each instance of X, and the result is presented in Table 1 and the graphs of Figure 7.The general algorithm used for the validation process is given in Algorithm 1.

SCM Ontological Framework Validation Results Interpretation
The results of the SCM validation are evinced in Table 1 and Figure 7, and section 4.3.1 explicates its meaning.

Understanding the Results
The LocalTest function, a Digitty library in R programming, is employed in the CIT validation of the results.Its takes the coordinates and the CIT criteria identified in SCM ontological framework of equation 6 or 6b as inputs alongside the dataset distribution to perform the CIT.The outputs of the LocalTest are the Pearson correlation coefficient estimates, the p-value, and the confidence interval (CI) of the correlation coefficient for every conditional independence that is assumed in the SCM DAG.The variation of the Pearson correlation coefficient is between -1 and 1.Thus with 0 output implies the presence of no correlation, and a -1 or 1 indicates a perfect linear correlation.The outputted p-value for the LocalTest function shows the probability of observing the distribution of the dataset under the assumption that the independence condition holds.Thus, if the experiment outputs a correlation coefficient of around zero, with a high p-value (>0.05), it is interpreted that the assumptions of the implied conditional independence in the model structure hold true.On the other hand, a high value for the correlation coefficient, with a small or low p-value, evinced that the conditional independencies assumed in the SCM DAG ontological framework do not hold in the dataset distribution.Further, the 2.5% and 97% columns show the correlational coefficient's 95% confidence interval (CI).Thus, the smaller or narrower the CI is, and the further away from 0, evinces a negation of the conditional independencies assumed in the SCM ontological framework.Notably, the CIT of a SCM DAG requires that the associate effects of the identified conditional (in)dependencies be zero.Therefore, each effect that is numerically zero or close to it affirms the CIT assumptions encoded in the SCM DAG ontological framework, and it is considered the outcome of the desire for the validation process.However, the reverse is the case when the outcome is statistically significant -shifting significantly away from the zero mark.When this happens, the validation process is not considered successful because there is something wrong with the SCM DAG ontological framework designed or the dataset distribution has issues [62]- [64], [93].
Another function of the Dagitty library tool used in R for performing the test is the plotLocalTestResults function.As the name implies, this function plots graphically the result of the LocalTest function, enabling the accurate visualization and interpretation of the output results.Therefore, when the plotted graph coordinates line points to the zero line, it signifies or validates the assumptions in the DAG -meaning there is no dependency or correlation between the variables in focus.And the further away from the zero lines, the stronger is the evidence against the implied CIT assumptions in the SCM DAG model [93].This function is also applied at a CI interval of 95% to show the uncertainty associated with the estimates.

SCM Ontology Results Interpretation
Thus, Table 1 and Figure 7 show the output results from the CIT, using the conditional independent assumptions of equation 6 or 6b, obtained from the SCM Direct Acyclic Graph (DAG) ontological framework of Figure 6.Since the variable set X in Figure 6 is only a representation of the 5 variables, which are gender, marital_status, state_Id, lga_id and Age variables, this CIT test uses the assumptions of equation 6, is performed 5 times for each instance of X, and the results is presented in Table 1 and Figure 7. From the results of the Pearson correlation coefficient estimates, it can be seen that all estimates of the coefficient are close to the zero mark, with a very narrow CI measured at 95%, and all p-values are above 0.05 (>0.05).Also, the plotLocalTestResults function shows the that the data distribution points are all align on the zero mark or closer to the zero.These results clearly validate the assumption evinced in equation 6 or 6b, which is concomitant to the validation if the SCM ontological framework for the admission process of Benpoly dataset.

LIME Explanations Presentation
This section presents the experiment result for LIME, where the LIME results produced fairness bias (section 4.41) and where the LIME fairness bias is constrained or ablated (section 4.4.2).

LIME Presentation Results with Fairness Bias
Figure 5 shows the dataset visualization after preprocessing and arriving at the nine ( 9) most important features in the admission process in the Polytechnic Education system in Nigeria.Eight of the nine features are considered input features, and one (Admission_status) is the outcome or targeted label.The prediction task is a binary classifier, where 1 depicts a prospective student who got admitted, and 0 depicts a prospective student who failed to be admitted.The black-box ML predictive algorithm employed is the Gaussian Naïve Based (GNB) classifier.The ML prediction accuracy is 89%, with an F1 score of 94%, and implementing the LIME explanation framework of Equations ( 2) and ( 3) with all the eight input features, the explanations results for the LIME individual feature weights as plotted in the graphs of Figure 8 for the eight inputs features are shown in Table 2. From the LIME explanations, the individual attributes features depicted as X in the SCM ontological framework of Figure 6 are known to have the highest weights (positively and negatively) that seem to influence the LIME explanations (i.e., X = {gender, martialstatus, stateId, lgaid, Age}).These LIME explanations for the Benpoly admission process (which is indeed an extrapolation for all admission processes in the Nigeria Polytechnic Education system) are obviously biased, as the results show that prospective students got admitted based on either their state of origin, sex, age, etc.These features are not the criteria for which a prospective student gains admission.Albeit, the LIME explanation framework made its explanation choices based on the patterns learned from the dataset (faithfulness or fidelity).Thus, these LIME explanations, as evinced, are not a true representation of the ground truth regarding the admission process in these institutions and, therefore, are not valid explanations for the admission process in the polytechnic admission process in Nigeria.These features, which are identified as having the highest weights, have nothing to do with the admission requirements in the polytechnic education system in Nigeria.Hence, there is a need for our fair-LIME explanation framework, which is implemented in the results of Table 3   The next narrative explicates the result outlined in Figure 8: Subplot 1: LIME Explanation for Instance 1 -Title: "LIME Explanation for Admission Subplot 5: LIME Explanation for Instance 5 -Title: "LIME Explanation for Admission Decision: Instance 5" -Description: "This subplot shows the LIME feature importance scores for the fifth instance in the test set.Unlike the previous instances, this applicant is predicted to be not admitted (AS = 0) with a probability of 1.00.The top contributing features are State of Origin (SO) and Local Government Area (LGA), but with negative importance scores, indicating that these factors are driving the non-admission decision.".Thus, from the explanations obtained from the automated LIME frameworks without our framework (i.e., constraining or ablating biased features), the LIME explanations are discovered to be biased in explaining important admission criteria.It only learned the patterns in the dataset, and made explanations on features such as state, gender, age, etc., which do not influence the admission selection criteria.

LIME Explanations Results with Ablated Input Features
To implement our fair-LIME explanation framework, the mutation of the SCM ontological framework of Figure 6 is imperative to perform the ablation process on the dataset.Thus, the five individual features identified and depicted as X, will need to be ablated from the SCM and concomitantly the dataset.This qualitative process is needed to identify, constrain, and sparse the features that can be inputted into the ML/LIME automated process to bring about predictions and LIME explanations with the dataset that are faithful and devoid of bias fairness bias.Hence, the fair-LIME SCM ontological framework is shown in Figure 9, as the X features are ablated or mutilated from the original SCM ontological framework of Figure 6.Thus, after the ablation process, our fair-LIME framework depicted in Figure 9, alongside constrained and sparse LIME Equations ( 4) and ( 5), will apply.Hence, the results for our fair-LIME framework results are shown in Figure 10 and Table 3.Therefore, instead of feeding the 8 input features into the automated LIME process, we now feed the LIME process with 3 input features (i.e., courseappliedid, modeofentry, and Current_Qualification) which are considered sine-quo-non for the admission process in Nigerian Polytechnic education system, as shown in Figure 9.The results from our fair-LIME framework results, as shown in Table 3 and Figure 10, has succeeded in removing the fairness bias in the initial LIME explanations results of Tables 2 and Figure 8.Thus, striking a balance between LIME explanations fidelity and fairness.Also, the black-box Gaussian Naïve based ML predictive model accuracy and F1 scores increased a bit to 90% and 95%, respectively.

Conclusions
This study successfully applied the KDD process to extract and discover knowledge from Benpoly's admission database in Nigeria.An application-based SCM ontological framework was designed to represent the admission process in the Nigerian polytechnic education system, leveraging the SCM's ability to identify causal relations among features and validate the ontology's correctness using the CIT criteria.
The SCM ontology was employed to identify features causing fairness bias in the automated ML predictions and LIME framework.By constraining and ablating these biased features, the proposed fair-LIME framework produced more stable and fair explanations compared to the original LIME framework, with improved prediction accuracy (91% vs. 89%) and F1 scores (95% vs. 94%).
The main contribution of this study is the novel integration of a qualitative SCM ontology with quantitative ML and LIME methods to enforce fairness in explanations.This approach addresses the problem of biased explanations generated by automated ML and LIME frameworks when based on patterns learned from datasets that violate fairness.The fair-LIME framework demonstrates the importance of incorporating domain knowledge through ontologies to identify and mitigate biases in ML explanations.
However, a limitation of the fair-LIME framework is its context-specificity, as the SCM ontology is derived from the background knowledge of a particular process.Therefore, the framework can only be extrapolated to similar contexts, and different contexts would require modeling new SCM ontologies for each process before applying the fair-LIME framework.
For future work, comparing the performance of other explanation frameworks, such as Shapley explanations, with the fair-LIME framework on the same dataset could provide valuable insights into their relative strengths and weaknesses.Additionally, exploring the application of the fair-LIME framework to other domains and datasets with different fairness challenges would further demonstrate its generalizability and potential impact on promoting fairness in ML explanations.

Figure 1 .
Figure 1.An image excerpt from the admission database in Benpoly.

Figure 2 .
Figure 2. Comprehensive ontological framework of the entire admission database, with classes, attributes, and their relations.

Figure 4 .
Figure 4. Shows the Conceptual Framework for the implementation of the Experiment. 1. Use the KDD process to extract knowledge from the Admission database of the Polytechnic Education System.2. Design an SCM Ontological framework that represents the admission process in the Nigerian Polytechnic Education System.3. Validate assumptions encoded in the SCM Ontological framework with the admission dataset (Benpoly dataset) using the CIT criteria.

Graph 4 :
Mode of Entry Distribution -Caption: "The bar graph illustrates the distribution of the Mode of Entry (ME) variable, indicating that most applicants apply through the JAMB route, while a smaller proportion apply through Direct Entry (DE)."-Annotation:"Mode of Entry distribution: 80% JAMB (ME = 1), 20% Direct Entry (ME = 0)".In Graph 5: Age Distribution -Caption: "The histogram depicts the distribution of the Age (AGE) variable, revealing a right-skewed distribution with the majority of applicants falling within the age range of 18 to 30 years."-Annotation:"Age distribution: Mean = 22.5 years, Standard Deviation = 3.7 years".

Figure 6 .
Figure 6.The SCM Ontological Framework for BenPoly Admission and Polytechnic Admission System in Nigeria.

Figure 8 .
Figure 8. Visualized LIME Graph plot for Local_Explanations for five data instances on All identified features in the SCM Ontology.Decision: Instance 1" -Description: "This subplot shows the LIME feature importance scores for the first instance in the test set.The applicant is predicted to be admitted (AS = 1) with a probability of 0.96.The top contributing features are State of Origin (SO), Local Government Area (LGA), and Mode of Entry (ME), while the other features have relatively smaller contributions."Subplot 2: LIME Explanation for Instance 2 -Title: "LIME Explanation for Admission Decision: Instance 2" -Description: "This subplot presents the LIME feature importance scores for the second instance in the test set.The applicant is predicted to be admitted (AS = 1) with a probability of 0.96.Similar to Instance 1, the top contributing features are State of Origin (SO), Local Government Area (LGA), and Mode of Entry (ME), indicating a consistent pattern in the model's decision-making process."Subplot 3: LIME Explanation for Instance 3 -Title: "LIME Explanation for Admission Decision: Instance 3" -Description: "This subplot displays the LIME feature importance scores for the third instance in the test set.The applicant is predicted to be admitted (AS = 1) with a probability of 0.92.The feature contributions follow a similar pattern to the previous instances, with State of Origin (SO), Local Government Area (LGA), and Mode of Entry (ME) having the highest impact on the prediction.".Subplot 4: LIME Explanation for Instance 4 -Title: "LIME Explanation for Admission Decision: Instance 4" -Description: "This subplot illustrates the LIME feature importance scores for the fourth instance in the test set.The applicant is predicted to be admitted (AS = 1) with a probability of 0.96.The feature contributions are consistent with the other admitted instances, highlighting the influence of demographic factors like State of Origin (SO) and Local Government Area (LGA) on the model's decision.".Subplot 5: LIME Explanation for Instance 5 -Title: "LIME Explanation for Admission Decision: Instance 5" -Description: "This subplot shows the LIME feature importance scores for the fifth instance in the test set.Unlike the previous instances, this applicant is predicted to be not admitted (AS = 0) with a probability of 1.00.The top contributing features are State of Origin (SO) and Local Government Area (LGA), but with negative importance scores, indicating that these factors are driving the non-admission decision.".Thus, from the explanations obtained from the automated LIME frameworks without our framework (i.e., constraining or ablating biased features), the LIME explanations are discovered to be biased in explaining important admission criteria.It only learned the patterns in the dataset, and

Figure 9 .Figure 10 .
Figure 9. Ablated SCM Ontological Framework for BenPoly Admission, where the bias features set X is mutilated from the rest of the SCM features

Table 1 .
Important variables and features used in the study.

Table 2 .
Changes in dataset size and composition during preprocessing.

Table 1 .
(6)ults of the CIT performed for each instance of X using Equation(6).
and Figures 9 and 10 in section 4.4.2.

Table 2 .
Results of the LIME explanations on ML predictions for 5 data instances with fairness bias.

Table 3 .
Result of the Ablation application on LIME explanations for 5 data instances