Data and Label Shift in LIME explanations
In this post, I will concisely summarize my research study, “A study of data and label shift in the LIME framework,” which was a collaboration with my supervisor, Professor Henrik Boström. The paper was accepted as oral in the Neurips 2019 workshop on “Human-Centric Machine Learning.” You can read the paper on Arxiv, and the workshop website can be accessed here: https://sites.google.com/view/hcml-2019. Introduction In 2019, LIME explanations were prevalent [1], but LIME operated differed significantly from how explanations functioned in the older days. Before LIME, feature importance explanations were the weights of an interpretable model, and they were one vector that provided importance scores for all instances. LIME operated differently and could provide feature importance for a single instance $x$. LIME calls this to be a local explanation. To obtain Local LIME explanations, we need a black-box prediction function that outputs probability scores, $f$, and a background dataset that is usually the training set $X$. So far, so good. But there are three steps where things start to get complicated: ...