ABSTRACT
This chapter dedicates to the techniques that help understand the way models process inputs into outputs. It adopts a tone which is factor-investing orientated and discusses examples related to machine learning (ML) models trained on a financial dataset. Quantitative tools that aim for interpretability of ML models are required to satisfy two simple conditions: that they provide information about the model and that they are highly comprehensible. In attempts to white-box complex machine learning models, one dichotomy stands out: Global models and Local models. Whereas global interpretations seek to assess the impact of features on the output overall, local methods try to quantify the behavior of the model on particular instances or the neighborhood thereof. Local Interpretable Model-Agnostic Explanations is a methodology originally proposed by Ribeiro et al. Their aim is to provide a faithful account of the model under two constraints: simple interpretability and local faithfulness.
