Abstract
We studied the utility of using machine learning algorithms in the estimation of feature importance and to visualize their dependence on Ethicality. Through our analysis and partial dependence plot we found linear relationships among variables and gained insight into features that might cause certain types of ethical behaviour.
Introduction
High-profile corporate scandals have underscored the critical need to understand the drivers of unethical conduct in professional environments. Traditional statistical methods, such as linear regression, have historically been used to link factors like gender, religiosity, and perceived pressure to ethical behavior. However, human psychology is rarely linear; traditional models often struggle to capture the complex, "messy" interactions between situational and demographic variables. This study leverages machine learning (ML) to bridge this gap, identifying key features and visualizing their non-linear impact on an individual's "Ethicality" score.
Methodology
The research utilizes three distinct ML techniques to evaluate feature importance and behavior:
- Support Vector Regression (SVR): Used for isolated feature importance (measuring $R^2$ for single features) and ablation studies ($\Delta R^*_k$) to determine how much the model's predictive power drops when a specific feature is removed.
- Random Forests (RF): Employed to rank features based on their predictive purity and frequency of use in decision tree nodes.
- Partial Dependence Analysis (PDA): Utilizing Gradient Boosting Regressors (GBR) to generate visualizations of the relationship between specific features and the target variable.
Data and Features
The study analyzed a dataset of 336 subjects from the Northeastern United States, comprising undergraduate students entering the workforce and currently employed graduate students. The dataset included 16 features, such as age, sex, marital status, education level, employment status, religiosity, income, and perceived pressure. The target variable, "Ethicality," is a numerical score representing ethical conduct.
Analysis and Insights: The Pressure-Income Dynamic
A core finding of this research is the distinction between linear and non-linear drivers of behavior. "Pressure" was identified as the primary driver of unethical behavior across all models. The relationship is predominantly linear: as perceived pressure increases, ethicality scores consistently decrease.
In contrast, "Income" exhibits a more complex, non-linear relationship with ethics. Partial dependence plots reveal a quasi-concave relationship, suggesting that predictability is highest at the extremes of the income spectrum. Individuals with very low or very high incomes show more distinct ethical patterns, while those in the middle-income range appear to be influenced by a more varied set of factors.
Key Results
The ML models provided a robust ranking of feature importance. In the isolated SVR analysis, "Pressure" was the only feature to achieve a significant $R^2$ (approximately 0.4), while other features showed negligible predictive power in isolation. The Random Forest model corroborated these findings, placing Pressure at the top of the importance hierarchy.
Ablation studies further highlighted the necessity of "Employment," "Education," and "Pressure" for model accuracy. Removing these features resulted in a significant drop in predictive performance, confirming their status as essential components of the ethicality model.
Conclusion
This study demonstrates that ML techniques can significantly enhance our understanding of ethical behavior by identifying non-linear interactions that traditional statistics might overlook. While "Pressure" remains the most significant linear predictor, the "Income Paradox" highlights the need for more nuanced models. Future research should expand the geographic scope beyond the Northeastern US and explore the causal relationships between the features themselves to further unravel the web of human ethics.
Summary updated