Intelligent Data Analytics Using Explainable and Interpretable Machine Learning Techniques

Abstract

With the widespread adoption of advanced machine-learning algorithms in data-driven decision systems, there is an increasing need for methods that accurately predict outcomes as well as provide understandable and transparent explanations. In the present research, an intelligent data-analytic framework was designed, in which the algorithm of machine learning was compared with each other and the machine learning explainability was assessed. To obtain repeated production of the synthetic binary classification data provided, a fixed random seed was used to generate a synthetic set of 500 observations, 6 continuous predictors and 1 outcome variable. The Logistic Regression, Decision Tree, Random Forest, Extreme Gradient Boosting and Neural Network models were tested using the accuracy, precision, recall, specificity and F1 score. Fidelity, stability, consistency, computational efficiency, and human interpretability were used for the assessment of Explanability, which involved the use of SHapley Additive exPlanations, Local Interpretable Model-Agnostic Explanations, Integrated Gradients, permutation importance, and partial-dependence analysis. The neural network had the best accuracy (86.2%), recall (86.7%) and F1 score (86.5%), while XGBoost had the best precision (86.8%) and specificity (86.5%). SHAP was the most explainable (4.36/5) and had good fidelity, consistency and interpretability. The feature-importance analysis indicated that the features 1, 2, and 3 are the most important features with values of 28.6%, 22.4%, and 18.8%, respectively. The analysis reveals that the quality of the explanations is one of the key indicators to consider in addition to the traditional indicators. These results, however, are synthetic (not empirical) and merely illustrative of the output of the analyses and not actual results. The framework should be further tested by using real and domain-specific data, multiple times of cross-validation, external populations, fairness evaluation and end-user evaluation. This validation may help to build generalizability, robustness and practicality of the use of the process in high-stakes decision-making contexts.