[Paper Review] A survey and taxonomy of methods interpreting random forest models

Maïssae Haddouchi, Abdelaziz Berrado|arXiv (Cornell University)|Jul 17, 2024

Forest ecology and management4 citations

TL;DR

This paper presents a comprehensive survey and taxonomy of interpretability methods for random forest (RF) models, categorizing existing techniques along multiple axes such as model-agnostic vs. model-specific, global vs. local interpretation, and feature importance estimation. It synthesizes current approaches to guide practitioners and researchers in selecting appropriate interpretation tools based on desired interpretability goals, offering a structured reference for improving transparency in RF-based machine learning applications.

ABSTRACT

The interpretability of random forest (RF) models is a research topic of growing interest in the machine learning (ML) community. In the state of the art, RF is considered a powerful learning ensemble given its predictive performance, flexibility, and ease of use. Furthermore, the inner process of the RF model is understandable because it uses an intuitive and intelligible approach for building the RF decision tree ensemble. However, the RF resulting model is regarded as a "black box" because of its numerous deep decision trees. Gaining visibility over the entire process that induces the final decisions by exploring each decision tree is complicated, if not impossible. This complexity limits the acceptance and implementation of RF models in several fields of application. Several papers have tackled the interpretation of RF models. This paper aims to provide an extensive review of methods used in the literature to interpret RF resulting models. We have analyzed these methods and classified them based on different axes. Although this review is not exhaustive, it provides a taxonomy of various techniques that should guide users in choosing the most appropriate tools for interpreting RF models, depending on the interpretability aspects sought. It should also be valuable for researchers who aim to focus their work on the interpretability of RF or ML black boxes in general.

Motivation & Objective

• To address the challenge of interpreting random forest models, which are often treated as 'black boxes' despite their inherent tree-based structure.
• To provide a systematic review of existing interpretation techniques in the literature for random forest models.
• To classify and organize interpretation methods along multiple dimensions, such as model-agnostic vs. model-specific and global vs. local interpretation.
• To support researchers and practitioners in selecting the most suitable interpretation method based on specific interpretability needs.
• To serve as a foundation for future work on interpretability in random forests and other machine learning black-box models.

Proposed method

• The authors conducted a comprehensive literature review of studies focused on interpreting random forest models.
• They classified interpretation methods based on key dimensions: model-agnostic vs. model-specific, global vs. local interpretation, and type of explanation (e.g., feature importance, counterfactuals, partial dependence).
• The taxonomy includes techniques such as permutation-based feature importance, partial dependence plots, individual conditional expectation (ICE), LIME, SHAP, and tree-based interpretability methods.
• The classification process involved analyzing methodological differences, interpretability objectives, and computational requirements.
• The paper organizes methods into a structured framework to enhance usability and guide selection based on application context.
• The survey emphasizes practical guidance and highlights trade-offs between interpretability depth, computational cost, and fidelity.

Experimental results

Research questions

• RQ1 What are the main categories of interpretation methods for random forest models, and how do they differ in approach and application?
• RQ2 How do model-agnostic and model-specific interpretation techniques compare in terms of accuracy, efficiency, and interpretability?
• RQ3 What are the strengths and limitations of global versus local interpretation methods in explaining random forest predictions?
• RQ4 Which interpretation techniques are most effective for identifying feature importance and understanding decision boundaries in random forests?
• RQ5 How can practitioners select the most appropriate interpretation method based on their specific interpretability goals and constraints?

Key findings

• The survey identifies multiple distinct categories of interpretation methods, including feature importance, partial dependence, individual conditional expectation, and post-hoc explanation techniques such as LIME and SHAP.
• Model-specific methods, such as permutation-based importance and feature contribution analysis, are shown to be more accurate and interpretable than model-agnostic alternatives in many cases.
• Global interpretation methods like partial dependence plots and individual conditional expectation provide insights into feature effects across the input space, though they may oversimplify complex interactions.
• Local explanation methods such as LIME and SHAP offer instance-level explanations but may vary in fidelity depending on the surrogate model and kernel choice.
• The taxonomy enables systematic comparison of methods, highlighting trade-offs between computational cost, interpretability, and fidelity to the original model.
• The study underscores the importance of method selection based on application context, with no single method being universally optimal.