Predicting Fatalities in Air Accidents using CHAID XG Boost Generalized Linear Model Neural Network and Ensemble Models of Machine Learning

Abstract

The study examines the historical data of about 4700 air crashes all over the world since the first recorded air crash of 1908. Given the immense impact on human beings as well as companies, the study aimed at utilizing Machine Learning principles for predicting fatalities. The train-test partition used was 75-25. Employing the IBM SPSS Modeler, the machine learning models used included CHAID model, Neural Network, Generalized Linear Model, XGBoost, Random Trees and the Ensemble model to predict fatalities in air crashes. The best results (90.6% accuracy) were achieved through Neural Network with one hidden layer. The results presented also include comparison of the predicted versus observed results for the test data.

Classification and grade prediction of aviation accident based on Lasso-SVR model

“…And the probability safety method is used to determine the possibility of the event, so as to evaluate the operability of ATC. This method effectively provides the correct decision basis for operational risk; Nikita [2] and others analyzed the fatality rate of aviation accidents caused by human factors by using CHAID, artificial neural network model and XGboost regression model. The effect accuracy of the model is 90.6%; Anushree H R et al [3] applied K-NN and SVM to weather type grouping, and predicted the level of aviation accident through probability selection tree; Jieli Huang [4] and others used STPA method to analyze the accident, and formed the aviation accident network by extracting the accident chain in order to improve the aviation safety benefit; Arjun h. Rao [5] and others developed a state based on accident model, which effectively removed the redundancy of NTSB coding system for flight out of control events; Chunyang Yang et al [6] used HFACS-BN model to code human factors to build BN model, and determined the most significant factors through the correlation estimation of conditional probability table; S. Koteeswaran et al [7] proposed the "improved oscillation correlation feature selection (IOCFS)" method, and compared with SVM, ANN, K-NN and multi class classifier, it is concluded that k-NN classifier has the best effect; Peter brooker [8] proposed a solution based on Bayesian belief network (BBN).…”

“…The evaluation indexes include training set score, testing set score, test set mean square error (MSE), test set mean absolute error (MAE) and2 And the comparison data of the evaluation indexes are shown in Table (3~6). Comparison data of SVR and Lasso-SVR evaluation indexes (Linear kernel function)…”

“…Table 6, it can be seen that Lasso-SVR mathematical model is obviously better than SVR mathematical model. The best optimization effect is to use Polynomial kernel function, its determination coefficient2 …”

Improving the Classification of Airplane Accidents Severity using Feature Selection, Extraction and Machine Learning Models

“…According to [13] airplane accidents can be caused by so many factors including: Pilot error due to miscommunication, distraction, exhaustion, drainage, etc. mechanical error, bad weather conditions, sabotage and human errors.…”

“…So many scientific contributions had tried to implement data0based approaches using machine learning (ML) models to deal with these safety issues specially in the context of our study which is the prediction of severity of airplane incidents. [4,6,7,12,13] propose ML-based models, deep learning are used and implemented on complex dataset that need deep models such as in [11,14]. Authors using machines learning (ML) models achieved promising results but there is always some data and implementations constraints including limited resources and information about fatal accidents because there are very rare to occur; that is why it is difficult to collect enough data to establish meaningful statistical analysis.…”