Evaluating decision trees as local explanation models in LIME

Details

Supervisors

Nijssen, Siegfried

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en science des données, orientation technologies de l'information, à finalité spécialisée

Abstract

enArtificial Intelligence (AI) and Machine Learning (ML) have grown significantly over the last decades. Neural networks for example, perform exceptionally well on many ML tasks. However, they are too complex for humans to understand easily, making it difficult to know why a specific prediction was made. They are considered as black-box models. This lack of transparency is a barrier to trust and accountability in AI applications. In the field of eXplainable Artificial Intelligence (XAI), many methods try to understand the predictions made by these black-box models. Local Interpretable Model Agnostic Explanations (LIME), is one of those widely used method for interpreting the predictions of complex ML models by providing local explanations, it was introduced by Ribeiro, Singh, and Guestrin[42]. LIME works by first generating new data points near an instance of interest, then each such generated data point is given as entry to the complex black box model to determine its prediction on the instance of interest. Next, a linear model is trained on the resulting training data generated points and finally, the coefficients of the features according to the linear model are used as explanation. Traditionally, LIME uses a linear model as Surrogate to provide explanations. This thesis investigates a variant of the LIME approach where we replace the linear surrogate model with a decision tree. We want to evaluate if decision trees can also serve as effective and interpretable local surrogate models for complex Machine Learning models explanation. To this end, we conducted experiments across seven benchmark tabular datasets using Random Forest classifiers as black-box models. Our contribution was to replace the linear surrogate model in LIME by a decision tree and to compare the explanations provided by both models on tabular datasets. Additionally, we investigated the issue of LIME in handling categorical features in tabular datasets that could make the explanations to be untrustworthy, we also wanted to see if with decision tree the explanations would be better as decision trees naturally handled them. Our results showed that decision trees can be effectively used as local surrogate models in LIME. They provided more intuitive explanations, since they explain with rules. But, they also showed weaknesses: they were more sensitive to LIME’s parameters like kernel width and neighborhood sampling, which sometimes made their explanations unstable. On the other hand, the linear surrogate was more consistent and reliable, even if sometimes less detailed. Both methods had difficulties when the dataset contained many categorical features.