Towards Enhanced Local Explainability of Random Forests: a Proximity-Based Approach
Trading and Investing / Artificial Intelligence & Machine Learning
Joshua Rosaler, Dhruv Desai, Bhaskarjit Sarmah, Dimitrios Vamvourellis, Deran Onay, Dhagash Mehta, Stefano Pasquali
The authors of the paper have proposed a novel apporach to a local explanability of the Random Forest (RF) algorithm!
Enhanced local explainability of random forests refers to methods and tools designed to shed light on the decision-making process of random forest models for specific instances or data points, as opposed to global explainability, which seeks to explain the model’s behavior as a whole.
Random forests, like many ensemble methods, are inherently more complex than simpler models, such as linear regression. While they can capture intricate relationships and deliver high accuracy, they lack transparency, making it challenging to interpret their decisions.
Local explainability aims to answer questions like “Why did the model make this specific prediction for this data point?”
The authors exploited the fact that the RF is nothing but an adaptive weighted K nearest-neighbors (KNN) model, i.e., it *learns* the distance metric as well as the number of nearest neighbors from the data (as opposed to imposing a distance metric and the number of nearest neighbors by hand as in the usual KNN algorithm).
Moreover, once these two objects become extracted from a trained RF for the given data. One can start explaining the out of sample prediction the same way as one would explain for the usual KNN algorithm.
This approach of explainability complements established methods like SHAP, which instead generates attributions for a model prediction across dimensions of the feature space whereas the proposed approach does so across the training data points.
Lastly, the authors implemented this approach in the context of a bond pricing model trained on US corporate bond trades, and compared it with various existing approaches to model explainability.
