Trustworthy mannequin explanations via energy-based conformal counterfactuals

Counterfactual explanations supply an intuitive and easy strategy to clarify opaque machine studying (ML) fashions. They work below the premise of perturbing inputs to attain a desired change within the predicted output.

If in case you have not heard about counterfactual explanations earlier than, be at liberty to additionally try my introductory posts: 1) Particular person Recourse for Black Field Fashions and a couple of) A brand new device for explainable AI.

There are usually some ways to attain this, in different phrases, many various counterfactuals might yield the identical desired consequence. A key problem for researchers has subsequently been to, firstly, outline sure fascinating traits of counterfactual explanations and, secondly, provide you with environment friendly methods to attain them.

Some of the necessary and studied traits of counterfactual explanations is ‘plausibility’: explanations ought to look reasonable to people. Plausibility is positively related to actionability, robustness (Artelt et al. 2021) and causal validity (Mahajan, Tan, and Sharma 2020). To attain plausibility, many present approaches depend on surrogate fashions. That is simple but it surely additionally convolutes issues additional: it basically reallocates the duty of studying believable explanations for the information from the mannequin itself to the surrogate.

In our AAAI 2024 paper, Trustworthy Mannequin Explanations via Power-Based mostly Conformal Counterfactuals (ECCCo), we suggest that we must always not solely search for explanations that please us however moderately concentrate on producing counterfactuals that faithfully clarify mannequin conduct. It seems that we are able to obtain each faithfulness and plausibility by relying solely on the mannequin itself, leveraging latest advances in energy-based modelling and conformal prediction. We help this declare via in depth empirical research and consider that ECCCo opens avenues for researchers and practitioners searching for instruments to higher distinguish reliable from unreliable fashions.

It is a companion publish to our latest AAAI 2024 paper co-authored with Mojtaba Farmanbar, Arie van Deursen and Cynthia C. S. Liem. The paper is a extra formal and detailed remedy of the subject and is obtainable right here. This publish is deliberately freed from technical particulars, maths or code. It’s meant to offer a high-level overview of the paper.