Explainable Supervised Learning Models for Aviation Predictions in Australia

1. Introduction

Recent years have seen a rapid proliferation of advanced artificial intelligence (AI) algorithms with profound performance regarding detection, prediction, and classification tasks [1,2,3,4]. Consequently, this has triggered an exponential rise in the development and deployment of AI-based applications across various industries for accelerated process automation. This change in basic assumptions has further been motivated by the evolution of Internet of things (IoT) technology which equips “things” that conventionally had no computing capabilities with some degree of intelligence through sensors and actuators that monitor the world around them, collecting data about every event and transmitting it to the concerned entities for further analysis and decision-making support [5]. For instance, the Airbus A350 comes with more than 250,000 sensors that monitor every component, in and out, of the aircraft and its surroundings and collect data ranging from weather to in-flight commands issued by pilots [6]. This translates into huge volumes of data flowing into air traffic management systems that require real-time analysis to support timely decisions by human stakeholders.

In non-safety critical fields, advanced data analytics techniques coupled with advanced AI algorithms have come to the rescue in such situations. However, the inherent lack of model transparency by AI models has turned out to be a great impediment to the deployment of AI in the aviation industry due to its sensitivity to safety-threatening occurrences [1,5,7]. The rationale behind AI models’ outputs remains opaque to human users to this day which is non-trivial to applications with high sensitivity to wrong model outputs, like the aviation industry, where even a seemingly slight misjudgment of the situation can have far-reaching catastrophic consequences ranging from loss of life to unrecoverable losses caused by accidents that often result in aircraft getting completely destroyed and aviation companies losing huge sums of money in compensating the victims’ families.

To achieve AI model transparency and trustworthiness, AI researchers are shifting their focus to finding approaches that can make models explainable both at a local level where the rationale behind individual model predictions is explained and at a global level which is concerned with giving an intuition of the decision boundary learned by the model while providing a high-level insight of how the learned features impact the model’s outputs [5]. This concept of making AI models explainable is known as Explainable Artificial Intelligence (XAI) [8,9]. Various XAI-based approaches have been proposed in the last half-decade including model-agnostic and model-independent approaches. Most of the traditional machine learning algorithms like support vector machines (SVM), Decision Trees (DT), and K-Nearest Neighbor (KNN), among others, are intrinsically explainable, that is, their mode of operation is easily explainable, and human users can easily visualize the rationale behind a given model output [8,10].

On the other hand, deep learning models continue to operate as a “Blackbox” with no human-centric explanations for their output owing to the enormous amount of non-linear mathematical operations that take place in their hidden layers [1,8,11,12]. Deep learning XAI approaches have been proposed to clear this barrier, including SHapley Additive exPlanations (SHAP) [13,14], DeepSHAP [13,14], Local Interpretable Model-agnostic Explanation (LIME) [15], DeepLIFT [16], among others.

In this paper, owing to the outstanding performance recorded by supervised machine learning algorithms in recent AI-based studies including healthcare [17,18], air transport [19], and computer vision [20] among others, and the labelled nature of the dataset used for this study supervised learning was deployed for all the experiments. We leveraged Variational Autoencoder (VAE) models for under-represented class instances augmentation to ease the class imbalance distribution of the ATSB dataset. Then we trained and evaluated four AI algorithms for a three-class classification problem. Out of the four models, three are standard machine learning algorithms (i.e., SVM, Logistic regression (LR), and Random Forest (RF)) while the fourth is a deep neural network (DNN) constituting five hidden layers. We then deployed the SHAP technique to each model intending to understand the features pertinent to the decision boundary of each model and how they differ [21,22].

Our contributions to this study are as follows.

• We investigate the impact of imbalanced distribution on the classification performance of various machine learning models and then propose a VAE-based data levelling approach to enhance model performance regarding minority class instances recognition.
• We evaluate and visualize the impact of each independent variable on the model output to allow model explainability and consequently make the model classifications more transparent and trustable by human stakeholders in the aviation industry.

The rest of this paper is structured as: Section II presents existing prior work related to this study followed by Section III which gives a detailed description of the various techniques deployed to bring this work to existence, Section IV presents the experimental results and comprehensive discussion of their implications. Finally, Section V concludes this paper with a brief account of the current state of research on XAI concerning aviation safety and the possible direction for future research as far as this topic is concerned.

2. Related Work

Despite capturing the attention of many AI researchers across various fields of research, Explainable Artificial Intelligence (XAI) is still in its infant stage, that is, not much has been done to fully explore it for most fields, aviation not being exceptional. This section presents the prior studies whose context was the deployment of XAI to the aviation industry.

A study by [22] presented a taxonomy of the various levels of XAI relating to the aviation industry that can enhance trust in machine intelligence to human users. The study also gives an insight into the different state-of-the-art XAI approaches highlighting the pros and cons of each and suggesting the most suitable application area for each. However, no quantitative-based experiments are conducted to prove their concept, and consequently, no quantified performance comparison of the approaches for real-world scenarios is done. The work in [23] proposed a framework for generating human-centric and scientific explanations to enable human users to describe and interpret model predictions in the aviation industry. The authors contextualized XAI at four levels including interactivity with users where the XAI provides interactive explanations, level of observation where XAI is realized by either studying the observed data or the observed model, model structure where explanations are either model-agnostic or specific, and purpose of the explanations.

In a case study, Midtfjord et al. [24] highlighted the complexities involved in accurately modelling the physical dynamics of runway surface friction incurred by aircraft during landing. This challenge arises due to the complex and nonlinear interactions among the various physical factors influencing surface friction, as well as their time-dependent interdependencies. By incorporating weather data and runway condition reports, the study employed eXtreme Gradient Boosting (XGBoost) models to develop a hybrid assessment framework capable of classifying slippery runway conditions and predicting slipperiness levels. To ensure model transparency and support reliable decision-making for airport operators, the SHAP (SHapley Additive exPlanations) method was applied, offering localized explanations of the model outputs.

By using data provided by the United States Air Force (USAF) spanning from 2010 to 2018, study [25] trained and assessed several versions of interpretable models based on Bayesian networks, designed to function as decision support systems for selecting USAF candidate pilots. The authors utilized SHAP values alongside split conformal prediction techniques to provide explanations for the resulting black-box model hence facilitating a transparent and manageable assessment of each feature’s impact on the model's prediction, as well as a reliable estimation of the prediction's uncertainty.

The study by Groot et al. [26] emphasized that in scenarios with high traffic density, analytical conflict resolution methods can lead to airspace instabilities. The authors also highlighted that, despite existing alternative AI-based methods, such as those leveraging deep reinforcement learning, having demonstrated promising results in these settings, they functioned as black-box models, making their decision-making processes challenging to interpret. To address this, the authors explored ways to explain the behavior of a Soft Actor-Critic model trained for vertical conflict resolution within a structured urban airspace. They employed a heat map of chosen actions as a tool to interpret and visualize the learning policy.

Study [27] introduced a green performance evaluation model for airport buildings, leveraging PCA and hierarchical clustering algorithms within an explainable, semi-supervised AI framework. This approach aimed to minimize reliance on human intervention when assessing green scores. By employing methods such as the scatter coefficient, Psi Index, variation, and permutation to reduce dataset dimensions, alongside divisive and agglomerative hierarchical clustering techniques, the authors were able to derive interpretable decision trees from the corresponding hierarchical dendrograms.

In their study, [28]developed a deep neural network model aimed at forecasting airport throughput based on both current and anticipated weather and air traffic conditions. To offer insights into which features the model relied upon to generate specific outputs, the authors proposed a method incorporating the explainable principal component analysis (EPCA) algorithm [29]. This approach enabled users to explore feature vectors derived from processing autoencoder outputs, providing a clearer understanding of the model’s internal processes and decision-making behaviour. Through the proposed method, users could gain insight into both what the model learned and how particular outcomes were achieved. The technique was applied to a real-world case study where deep learning was used to predict the total number of flights—either landings or takeoffs at major US airports within a 15-minute interval.

Zeldam et al., [30] proposed and applied an automated failure diagnosis model where feasibility was tested using XAI on the RNLAF (Royal Netherlands Air Force) F-16s dataset. Their approach was demonstrated using several learning models---Support Vector Machine (SVM), Random Forest (RF), Naive Bayes (NB), and Neural Network (NN)---and experimental findings showed RF to record a higher performance of ≃ 80%. Their research is also based on XAI for model interpretation and understanding of the assessed diagnosis. The survey conducted in [5] revealed that LR, SVM, RF, and NN models are better in AI prediction and hence can be used on aviation datasets. Their survey was on AI and XAI in Air Traffic Management (ATM) where they identified the current and future research developments in the same domain. They indicated that LIME and SHAP explanations are now commonly used as they are acceptable to researchers and thus should be used to give more knowledge and understanding to what models give to avoid biases and trust issues among the end users. They also studied the importance of using autoencoders in AI models as explained by Dubot and Olive in their studies specifically regarding air transportation [31,32].

In [33] a CNN (Convolutional Neural Networks) model was developed for the detection of topological defects in automated fiber placement manufacturing. To gain an intuition of the visual features (individual pixels or a group of pixels forming small image segments) that contribute to the model’s decision boundary, three XAI techniques were deployed including Smoothed Integrated Gradients (SIG), Guided Gradient Class activation mapping while the importance of deep learning features was obtained using Shapley Additive Explanations. For visualization of model classifications, SIG and DeepSHAP techniques were deployed. The authors computed and visualized explanations for sample layup defective images while seeing how these explanations correlated with model decisions. Also, study [34] developed an inception network model (CNN-based model) to discriminate between classes of objects/images. For model explainability, the authors deployed image feature visualization techniques including, visualizing activated neurons, deconvolution, deep dream, and Mask R-CNN segmentation to the output of each layer to extract intermediate information and understand which features are learned for input image discrimination. The authors in [33] presented a CNN-based semi-supervised model that used an autoencoder component for unsupervised learning and a fully connected network component for supervised learning. The study leveraged Linear adversarial perturbation using the fast gradient sign method (FGSM) to realize input feature perturbation in a bid to understand the impact of each feature on the model classification performance. Also, t-SNE was initialized using PCA (Principal Components Analysis) and then deployed to visualize the high-dimensional latent feature space in a 2D space.

In [35] a recurrent LSTM-based deep neural network was trained and evaluated on a historic flight dataset collected at the John F. Kennedy International Airport for predicting future unsafe situations. To ensure transparency of model operations, the authors leveraged a combination of input perturbation techniques and an intrinsically explainable hyperplane-based classification model to, in a global approach, explain the decision boundary of the proposed neural network model. The proposed LSTM (Long Short-Term Memory) model recorded 9.4%,12.8%, and 42.3 seconds while the hyperplane-based model recorded 97.2%, 18.0%, and 51.8 seconds in terms of accuracy, miss-detection rate, and duration before the occurrence of the degraded state, respectively.

In a study [36], Saraf et al. proposed a prototype tool for verifying and validating AI-based aviation systems. A LIME-based technique was used to generate reliable, human-understandable explanations for predictions from the Multiple Kernel Anomaly Detection (MKAD) algorithm, an SMV-based algorithm developed by NASA (National Aeronautics & Space Administration) for aircraft trajectory anomaly detection. Using the enhanced input perturbation technique, the proposed approach located exact anomaly points from the aircraft trajectory images. The authors also proposed a prototype for verifying and validating AI-based aviation systems where they applied XAI for aviation safety applications. Their work gave proof of concept for an explanation as they used XAI to make reasonable interpretations that humans can understand in making decisions based on the models. Hence concluded that the application of XAI can be used by subject matter experts (SMEs) and is useful concerning users accepting the decision made by AI-based support tools.

In the study [37], a genetic algorithm-based conflict resolution algorithm for ATM was proposed to determine the best aircraft trajectories. Model explainability was evaluated by assessing the degree to which the controller understood its operations. Flight trajectory data was collected via flight simulations with nine controllers and the authors designed questionnaires to get feedback from each controller. To get fine-grained details, the authors also designed and utilized semi-structured questionnaires to assess the explanations at four levels, that is, at a Black box level where only the selected solution was presented, at a heat map level where a combination of potential solutions was demonstrated and at a storytelling level where human-centric explanations in form of storytelling were used.

Despite many of the above studies emphasizing the importance of working with explainable and interpretable models in safety-critical fields, no prior studies have explored explainable models in the classification of aviation incidents using the ATSB aviation incident reports, while the few that have studied aviation incidents using this dataset have not considered the impact of its imbalanced distribution on model performance. This study closes this knowledge gap by training and evaluating the performance of various classic machine learning models and a deep learning model in classifying aviation incidents in the Australian Aviation industry while giving an insight into how the various features influence model predictions for both balanced and imbalanced learning.

3. Materials and Methods

The experimental execution in this study followed a three-step approach starting with data preprocessing followed by model training and evaluation and finally, the application of XAI-based techniques, SHAP and LIME, for model explainability. To study the impact of class imbalance, the investigated models were trained and evaluated on both the imbalanced and balanced datasets before and after the application of VAE.

3.1. Dataset

This study considered Accidents, Incidents, and Serious Incidents that were recorded in Australia for 20 years resulting in a dataset with 26,262 records distributed among the three classes as 3,402 (12.95%) “Accidents”, 20,894 (79.56%) “Incidents”, 1,966 (7.49%) “Serious Incidents” where the data was sourced directly from the ATSB investigation authorities.

3.2. Data Preprocessing

To ensure the dataset was in the desired format various data analytics techniques were utilized for the identification and removal of outliers and missing values, elimination of irrelevant features, handling of class imbalance, and feature encoding [36] Categorical features including “State”, “FuelType”, “RegistrationType”, “ActivityType”, and “ActivitySubType” were encoded as numerical features using 0ne-hot encoding. To handle class imbalance, additional instances were generated for the two under-represented classes— the “Accidents” class and the “Serious incidents” class—using a variational auto-encoder (VAE) [38]. The use of VAEs for imbalanced learning in this study was inspired by the impressive performance recorded by VAEs in enhancing the recognition of minority class instances for intrusion detection in the study [39]. A detailed description of how VAE operates is presented in the Subsection below.

3.3. Variational Autoencoder for class imbalance

As seen from the previous subsection, the dataset distribution in terms of the “occurrenceClass” variable is uneven with the incident class instances forming almost 80% of the total instances while accident and serious incident instances share the remaining 20%. Learning from imbalanced datasets can result in biased learning whereby the model learns patterns of the majority class and nothing from the under-represented classes. In such a case accuracy can be misleading since it can be high even when no minority class is assigned to the right class. This can be seen in section IV where we present validation results of models trained on the imbalanced dataset. To mitigate this issue, we employed variational autoencoders to generate more instances for the two under-represented classes.

A variation autoencoder first introduced [38], is an artificial neural network-based generative model that operates by learning a set of parameters, $\varnothing ,$ of an approximate posterior, $q_{\varnothing }\left(z/x\right)$ for the unknown true posterior (encoder),$p_{\theta }\left(z/x\right)$, where $x$ is a random variable from the unknown underlying prior dataset distribution, $p_{\theta }\left(x\right)$, of the training dataset while $z~N\left(\mu ,\sigma \right)$ is the low-dimension latent space of the learned VAE and often approximates a normal distribution. The decoder, $p_{\theta }\left(x/z\right)$, is trained to reconstruct the input variable, $x\text{'}$, by sampling from the learned latent space distribution, $zz$ so that $x\cong x\text{'}$. The model learns by minimizing two errors, i.e, the Kullback Leibler (KL) divergence between the actual and approximate posterior and the reconstruction error between the input data points, $x,$ and reconstructed data points, $x\text{'}$. Figure 1 depicts the general architecture of a VAE.

In this work, we developed two VAE models for the two under-represented occurrence class instances of the ATSB dataset. For instance, to generate more samples for the serious incident class, we trained a VAE using only samples of the serious class and then used the decoder model of the trained VAE to generate additional samples for that class. The same procedure was followed to generate additional “accident” class instances to balance the distribution between the three classes of instances of the “occurrence” class variables. 16000 and 17000 instances were generated for the “Accident” and “Serious Incident” classes respectively.

3.4. Model Performance Evaluation

This section presents a description of the evaluation metrics that were used in this study to assess the performance of the model. This work focused on multi-class classification and measured the performance by considering how well the predictions were distributed among classes based on precision, recall, F1 score, and accuracy [40,41,42,43,44]. F1-score as one of the performance metrics used in this study is simply the weighted combination of both recall and precision and its range is between 0 and 1 and the higher value is more preferred [41]. Table 1 shows the equations and explains how these performance metrics are calculated.

Where.

TP denotes the True Positive value where the actual value is similar/equal to the predicted value (i.e; the value of accident predictions that is correct according to the labeled data).

TN: The True Negative value for a class where it is the sum of all columns and rows except the class's values being calculated.

FP: False Positive value where the class is the sum of values of the corresponding column except for the True Positive value (i.e; the value of accident predictions which is not correct according to the labeled data).

FN: False Negative value where the class is equal to the number of wrongly predicted values or not detected by the model (i.e; the value of accident labeled in the original data but not/wrongly detected by the model).

3.4.1. Confusion Matrix

A confusion matrix is a square matrix in the R^m dimensional space, where m is the number of unique entries of the dependent variable instances distributed among the class labels by the trained AI model during the testing phase. It is a tool used to give visual information of how well the model performed on the test dataset and is often used as a yardstick to measure the goodness of the model. Table 2 demonstrates a simple confusion matrix for a binary classification problem discriminating between positive and negative instances of the test dataset. The dark cells show the correctly predicted value i.e., true positives (TP) and true negatives (TN), while the lighter cells show the wrong predictions i.e., false negatives (FN) and false positives (FP) [43].

3.5. Model Training and Validation

As indicated earlier, four classifiers including Logistic Regression, (LR), Random Forest (RF), Support Vector Machine (SVM), and the conventional feedforward Deep Neural Network (DNN) with backpropagation, were trained and evaluated for discriminating between instances of the occurrence class of the ATSB dataset. For both imbalanced and balanced dataset scenarios, python’s scikit-learn library was utilized to randomly split the datasets into train and test sets in a 4:1 ratio. The same library was used for instantiating the three standard machine learning classifiers (LR, RF, and SVM). It is worth noting that the three models’ hyper-parameters were set to sklearn’s defaults except the max_ depth on RF, which was set to 30, and the SVM kernel which was set to linear.

Also, we maintained the common architecture for the DNN which conventionally constitutes an input layer followed by a finite number of hidden dense layers (five hidden layers for the DNN in this study) and an output layer. For all hidden layers, a Rectified Linear Unit (ReLU) was deployed for activation while the output layer deployed softmax as the activation function. To prevent the model from overfitting, each of the first three hidden layers is followed by a dropout layer that in a stochastic manner selects 20% of nodes in each weight cycle [45]. To generate the predicted class, the argmax function was used. This returns the index corresponding to the entry with the highest probability from the softmax output.

Python libraries including Shap, pandas, and Numpy were respectively used for model explainability, loading and managing dataframes, numerical dataset transformation and label encoding while Matplotlib and Seaborn served to generate visual plots of model scores. All experiments were implemented in a Jupyter notebook on a Linux server with 256 CPU cores and 256GB RAM running Ubuntu OS.

3.6. Explanation Methods

To assert model transparency and trust by human users, we deployed XAI to generate model explanations. XAI is a technique applied in AI such that the results of a specific decision can be understood by humans. Various algorithms have been proposed in the recent past for the generation of individual prediction-based explanations, also known as local explanations. Local Interpretable Model-agnostic Explanation (LIME) proposed by [15] is a good example of such a technique which has seen application in numerous studies [46,47]. However, LIME and similar models have the drawback of the inability to explain the model’s predictions at a global scale, therefore, we utilized the SHAP module to generate and visualize global explanations for each of the classifiers used in this work. Despite the structures and prediction processes of the three classic ML (Machine Learning) algorithms (LR, RF, and SVM) being inherently explainable, we still applied SHAP to them to allow for a visual comparison between the features they learn and those learned by the DNN to define the decision boundary.

3.7. SHapley Additive exPlanations (SHAP)

This technique was proposed by [14] as a way of transforming “black box” AI models into transparent “white box” models to enhance their trustworthiness by generating and visualizing global explanations for the model’s learned decision boundary. SHAP was proposed to enable the explainability of complex AI models including ensembles and deep neural networks which were hard for even AI experts to explain. It works by assigning a numerical value (known as the shap_value) to each feature in the train set that defines its degree of importance to the model’s outputs. SHAP deploys the feature importance additive measure to generate a unified solution that combines the model explanation benefits of the previous local explanation XAI-based schemes including LIME [15], DeepLIFT [16,48] Layer-wise relevance propagation [49] and Classic Shapley Value Estimation [50,51,52], to generate global explanations. Figure 2 gives a high-level visual overview of the scope of XAI deployment in this work.

4.1. Impact of Dataset Balancing on Model Performance

The results demonstrate that balancing datasets using Variational Autoencoder (VAE) significantly improves classification performance across all models (SVM, RF, LR, DNN). Minority class recognition, particularly for accidents and serious incidents, showed marked improvements, with F1 scores increasing from 0.0 on imbalanced datasets to nearly 1.0. These findings reaffirm the necessity of robust preprocessing methods like VAE in addressing class imbalance—a persistent challenge in aviation safety datasets. However, the results also expose the limitations of relying solely on accuracy, as models can be overfit to majority classes. Confusion matrix analyses further support the use of metrics like precision, recall, and F1 scores for evaluating model performance in safety-critical contexts.

4.2. Explainability and Feature Importance Analysis

Using SHAP, the study provided insights into feature importance, revealing that models trained on balanced datasets relied more on domain-relevant features such as Latitude, Longitude, and Activity Type. In contrast, imbalanced datasets led to weaker feature selection and poor minority class recognition. The DNN model, while leveraging nuanced relationships, remains inherently opaque, highlighting a critical gap in trust and transparency for stakeholders.

4.3. Challenges in Transparency and Trust

Although SHAP adds interpretability to AI models, its scope is limited to local and feature-level insights. Deep learning models like DNNs, despite their high performance, pose challenges for adoption in aviation safety due to their "black box" nature. Developing hybrid explainability techniques that offer comprehensive global and local insights remains a vital area for future research.

4.4. Practical Implications and Future Work

The study underscores the importance of balanced datasets and XAI tools for improving model acceptance in the aviation industry. Collaborating with domain experts to align AI systems with operational workflows is critical for fostering trust. Future research should extend these findings to diverse datasets, explore hybrid models, and develop aviation-specific explainability tools to further bridge the gap between AI performance and stakeholder trust.

5. Conclusions

This study addresses two critical barriers to adopting AI in aviation safety: class imbalance and explainability. Through VAE-based dataset balancing, substantial improvements in recognizing minority class instances were achieved, underscoring the importance of addressing class imbalance in safety-critical applications. Additionally, the integration of SHAP enhanced model transparency by providing actionable insights into feature importance, which are essential for building trust among stakeholders. These findings emphasize the need for continued innovation in hybrid explainability frameworks that combine the predictive power of deep learning with interpretability. By offering practical strategies for balancing datasets and enhancing model transparency, this research lays the foundation for a safer integration of AI in aviation safety workflows. Future work should expand these findings to other datasets and incident types while developing domain-specific explainability tools to support broader AI adoption in the aviation industry.