ETD

• Behavioral Equialence
• Decision Trees
• Exploratory Analysis
• Interpretable Models
• KNN
• Logistic Regressors
• Model Classification
• Model Equivalence
• Model fairness
• Rashomon Effect
• Rashomon Set
• Supervised ML

The Rashomon Effect describes a phenomenon in machine learning where multiple models achieve nearly identical predictive performance on the same data while using different internal logics. While model selection typically relies on accuracy, the existence of a Rashomon Set (RS) enables the inclusion of additional evaluation criteria such as fairness, reliability, and interpretability. This thesis proposes an inter-family, multi-dimensional framework to define and identify Functional Behavioral Equivalence among interpretable models. The research examines whether models within the RS exhibit similar behaviors beyond performance. We focus on Decision Trees, Linear Regressors, and K-Nearest Neighbors, applied to four tabular classification tasks in sensitive domains like finance and law enforcement. The methodology includes training a diverse pool of models, constructing the RS based on generalization, and performing exploratory equivalence analysis via clustering. Results provide a map of interchangeable solutions, supporting the selection of models that are not only accurate but also ethically and behaviorally aligned.