One Explanation Does Not Fit All: The Promise of Interactive Explanations for Machine Learning Transparency

Abstract

The need for transparency of predictive systems based on Machine Learning algorithms arises as a consequence of their ever-increasing proliferation in the industry. Whenever black-box algorithmic predictions influence human affairs, the inner workings of these algorithms should be scrutinised and their decisions explained to the relevant stakeholders, including the system engineers, the system's operators and the individuals whose case is being decided. While a variety of interpretability and explainability methods is available, none of them is a panacea that can satisfy all diverse expectations and competing objectives that might be required by the parties involved. We address this challenge in this paper by discussing the promises of Interactive Machine Learning for improved transparency of black-box systems using the example of contrastive explanations -a state-of-the-art approach to Interpretable Machine Learning.Specifically, we show how to personalise counterfactual explanations by interactively adjusting their conditional statements and extract additional explanations by asking follow-up "What if?" questions. Our experience in building, deploying and presenting this type of system allowed us to list desired properties as well as potential limitations, which can be used to guide the development of interactive explainers. While customising the medium of interaction, i.e., the user interface comprising of various communication channels, may give an

On Selective, Mutable and Dialogic XAI: a Review of What Users Say about Different Types of Interactive Explanations

“…Work pertaining to the technical aspects of XAI also identify the importance of such "user-centric" explanations. [100,108]. Numerous papers have emphasized the need for explanations that are tailored to the context, audience and purpose of the explanation [2,32,33,36,90].…”

“…For each paper in their corpus, they documented personalized explanation properties (complexity, content and presentation), personalization granularity (to each user or per category of user) and personalization automation (manual or automatic). Additionally, they observed that personalization of explanations can be either iterative or one-off, with user information being collected once prior to showing explanations [100,108]. While the personalization of explanations is particularly important given the role of explanations in filling one's specific knowledge gaps, we believe there is a greater granularity of interaction to explore beyond the categories mentioned in [100].…”

“…These efforts have resulted in numerous examples of sophisticated interactive interfaces integrating sometimes complex XAI techniques. For example, the strand of research called "conversational XAI" made significant strides in providing explanations in natural language to a wide range of user questions [43,44,108].…”

Influence of context on users’ views about explanations for decision-tree predictions

“…Cheng et al (2019) found that their interactive interface, which allowed users to modify the value of features and see the impact of this change on the prediction of a linear regressor (what if), increased users' objective and self-reported understanding of the ML model compared to a static interface, which did not allow such changes. Sokol and Flach (2020b) studied counterfactual explanations for DTs in an interactive system where users could change or remove features, or request an explanation for a hypothetical instance. Counterfactual explanations were generated by representing a tree structure as binary meta-features, and selecting the shortest statement that minimizes an L1-like metric compared to the DT trace.…”

“…Reiter (2019) argued that good explanations must be written for a specific purpose and audience, have a narrative structure, and use vague language to communicate uncertainty. The explanations generated in (Sokol and Flach, 2020b) and (Biran and McKeown, 2017) have a narrative structure, and those in (Biran and McKeown, 2017) use vague language to convey strength of evidence. A different perspective is offered by expectation theory, which posits that the surprisingness of an event may stem from a discrepancy between the state of the world and propositions that are deducible from presented information (Ortony and Partridge, 1987).…”