Recent works on post-hoc example-based eXplainable AI (XAI) methods have focused on counterfactual explanations to provide justifications for predictions made by AI systems. Counterfactuals explain by showing what changes to input-features change the output decision. However, a lesser-known, special-case of the counterfacual is the semi-factual, which provide explanations about what changes to the input-features do not change the output decision. Despite their significant potential, semi-factuals have largely been unexplored in the XAI literature. My PhD research aims to address this gap by establishing a comprehensive framework for the use of semi-factuals in XAI. This includes development of novel methods for their computation, validated through user studies.
In recent years, research on eXplainable AI (XAI) have garnered significant attention as it aims to
improve the transparency and interpretability of the black-box AI models. Among diferent XAI
strategies, post-hoc example-based explanation methods which provides after-the-fact justification
have been very popular. Within this landscape, significant efort have been expended on counterfactual
explanations [
Counterfactual explanations inform users about how the output-decision can be altered by changing
key input-features in the form of "if only" reasoning. For example, when a customer is refused a loan, the
counterfacutal might say "if only you asked for a loan with shorter term, it would have been approved".
Semi-factual explanations, on the contrary, inform users about how the output-decision remains the
same when the key input-features change using "even if" reasoning. So, in the banking recourse example,
the semi-factual might say "even if you doubled your income, your loan would still be refused". In
cognitive science, counterfactuals and semi-factuals have been shown to have diferent psychological
impacts on users where the former tends to enable strong causal relation whereas the latter weakens
the causal support [
The origin of semi-factual as explanations can be traced back to early works on post-hoc explanatory
case-based reasoning (XCBR)[
The aim of my PhD thesis primarily focus on defining the characterstics of semi-factual explanations and establishing a comprehensive framework for their use in XAI.
My PhD research seek to establish a formal definition for semi-factuals in XAI unifying their computational (i.e, "what needs to be computed") and cognitive requirements (i.e, "the response to be elicited in users"). Based on the desiderata, novel methods will be developed to generate these explanations. This work also intend to propose key evaluation measures to asses their quality and hence provide a strong foundation for future advancements in this area. Furthermore, user studies will also be conducted to analyze how people perceive such explanations.
My dissertation focuses on 3 main research questions: • RQ1: What does the prior literature on semi-factuals in the field of Cognitive Science, CBR and AI tell us about the fundamental characteristics and desiderata for their use? The aim is to systematically identify and articulate a set of desirable attributes for the utilization of semi-factual explanations in the context of AI. • RQ2: What novel methods can be devised to efectively generate, interpret and evaluate semifactual explanations in AI systems? This research question addresses the core objective of my PhD thesis which is to explore innovative methodologies for the generation of semi-factual explanations that meet the established desiderata. • RQ3: How do people comprehend and interpret semi-factuals and how do these explanations impact their trust and understanding of AI systems? The objective is to conduct comprehensive user tests to analyze their impact and efect on people.
The progress so far can be divided into three phases.
In my first phase of work [
In cognitive science, semi-factuals have been extensively studied in Philosophy [
Similarly, a number of methods were proposed in early CBR research where semi-factuals were
characterized as a fortiori arguments. Doyle et. al [
The proposed novel MDN method finds the furthest neighbor of the query along some featuredimension while being in the same class as query (this is analogous to the use of NUNs in counterfactuals, where an existing datapoint is used as an explanation). MDNs are known data-points in the dataset that share some common features with the query but are far from it on some key-feature.
To compute MDNs, for a given feature of q, its neighbours along that dimension are partitioned into instance-sets that have higher values (i.e., HighSet) or lower values (i.e., LowSet) than the query. Each of these sets are ranked-ordered separately using a Semi-Factual Scoring (sfs) function, which is a distance measure that prioritises instances that are sparse (few feature diferences) while also having the highest value-diferences on the selected dimension, as follows: sfs(, , ) = where S is High or Low Set and ∈ , same() counts the features that are equal between q and x, F is the total number of features, dif() gives the diference-value of key-feature, f, and dif () is the maximum diference-value for that key-feature in the HighSet/LowSet. Basically, the instance with the highest overall sfs value from the HighSet/LowSet is the best MDN for that feature. This computation is done for each feature of q, independently, with the best of the best instance (i.e., with the highest sfs value across all features) is selected to be the semi-factual for the query as shown in Figure 1.
SFMDN(, ) = arg max () ∈
We show that MDNs meet many of the desiderata for semi-factuals though they may not be an
optimal solution. Furthermore, we experimentally compared four historical CBR methods (three
KLEORvariants [
In the second phase, a series of experiments were performed to propose two new MDN variants to overcome their initial limitations. The custom sfs() function in MDN scores each candidate instance based on their relative feature-value distance as well as closeness to the query. Along this line, we modify the scoring function to optimize their behaviour. (1) (2)
The two components in the scoring function are equally weighted to find the best semi-factual. Since original MDN (MDNv1) performed relatively poorly in the sparsity metric we propose SparseMDNs (MDNv2), which prioritises the similarity between non-key features. Essentially, we introduce a regularizer in the original sfs() function, which penalizes the algorithm for finding semi-factuals with higher feature-diferences, thus promoting sparse explanations. We modify the scoring function by weighing it with the proportion of features that are "not same" between the query and the instance. Hence, instances with higher number of similar features will be assigned high scores to obtain sparse MDNs.
1 dif ( , ) sfs2(, , ) = − (, ) * ( dif ( , ) +
In both MDNv1 and MDNv2, the similarity between non-key features between query and instances using same() involves a direct comparison of their values. Specifically, the function checks if the values are identical in case of categorical features, while the continuous features are considered same if they fall within a predefined threshold range. However, it is not always straight-forward to determine the optimal threshold and it may vary across diferent features. Hence, we propose Dist-MDNs (MDNv3) where we modify the scoring function to compute similarity directly in the feature space as: sfs3(, ) = dif ( , ) * dist( , ) where dist() computes the 2-norm distance and q and x represents the query and instance with only non-key features (i.e excluding the key-feature in consideration) respectively.
We also performed comprehensive tests to evaluate the proposed methods. The results show that the proposed variants could improve on the initial MDN limitations, however, the historical CBR methods still perform better on some measures.
Finally, in my most recent work, comprehensive tests were conducted to determine if counterfactuals are needed to obtain the best semi-factuals.
We divide the literature on semi-factual methods into two groups: Counterfactual-Guided and Counterfactual-Free. The Counterfactual-Guided group uses counterfactuals as guides to find the semi-factuals, whereas, Counterfactual-Free methods considers exploration within the query-class without explicitly relying on counterfactuals. Specifically, we consider the key question “Are the best semi-factuals found by using counterfactuals as guides?”. We evaluated 8 semi-factuals methods (4 from each group) on five evaluation metrics that attempt to capture key aspects of the desiderata for “good” semi-factuals. The results indicate that counterfactual-guidance is "not" necessary to find best semi-factuals. Each method do well on one or two metrics but then poorly on others. However, there is no single method that is consistently good across all measures. This work has been accepted in the main proceedings of ICCBR’24. (4)
Overall, [RQ1] have been successfully addressed and completed, and [RQ2] is well underway with an in-depth exploration. Based on the findings so far, the current focus is to devise novel generative methods to obtain semi-factual explanations. In the future, a primary emphasis will be on conducting user tests. The objective will be to study human perception of these explanations and gather insights about their impact, particularly within the context of AI systems. The collected user feedback will contribute to a comprehensive understanding of the nature and utility of semi-factual explanations in AI, thereby facilitating further advancements in this field. The results of these studies will address the objectives of [RQ3].