Posts

Note: This blog post summarizes my research paper, “The Blame Problem in Evaluating Local Explanations and How to Tackle It”. The article is available on [Arxiv] (https://arxiv.org/abs/2310.03466) and was accepted in ECAI XAI^3 Workshop. The workshop website can be accessed here. TLDR; The evaluation measures of local explanations suffer from a problem called blame problem. The blame problem is when we blame the explanation wrongfully for their unfaithful explanations when the fault can be traced back to the black-box model as the oracle. The blame problem can be avoided when ground truth feature importance scores are directly extracted from the model we aim to explain. ...

In this blog post, I would like to provide a summary of my research paper, “Can Local Additive Explanations Explain Linear Additive Models.” The paper was accepted at the journal track of the European Conference on Machine Learning 2023. The list of all accepted papers can be seen here. You can access the paper in the Data Mining and Knowledge Discovery journal. TL;DR We proposed an evaluation measure for local additive explanations called Model-Intrinsic Additive Score (MIAS). Using MIAS, we showed that linear additive explanations such as LIME and KernelSHAP can explain Linear Regression models faithfully. On the other hand, they fail to provide faithful explanations when explaining Logistic Regression and Gaussian Naive Bayes. ...

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: ...