TL;DR
This paper introduces "semi-factual" explanations as a crucial, yet under-explored, aspect of Explainable AI (XAI). It surveys existing work, defines key desiderata for semi-factual XAI, and presents a benchmark study comparing historical and a novel method to establish a foundation for future research in this area.
What Problem Does This Paper Address?
The field of Explainable AI (XAI) is increasingly focused on "counterfactual explanations," which tell users how to change input features to achieve a desired outcome (e.g., "if you had a shorter loan term, your loan would have been approved"). However, this paper highlights a related but distinct type of explanation: "semi-factuals." Semi-factuals explain what changes to input features would not alter the outcome (e.g., "even if you doubled your income, your loan would still be refused"). While these "even if" explanations are psychologically distinct and potentially valuable for understanding system limitations and causal relationships, they have received significantly less attention in AI research compared to their counterfactual counterparts. This paper aims to bridge this gap by systematically reviewing the literature, defining requirements for semi-factual XAI, and providing empirical benchmarks.
Methods
The authors conducted a systematic survey of literature on semi-factuals, drawing from philosophy, psychology, and AI. They then defined a set of desiderata for semi-factual XAI, encompassing both computational requirements (what needs to be computed) and psychological requirements (the desired user response). Based on this, they implemented and benchmarked five methods:
- Three KLEOR variants (Sim-Miss, Global-Sim, Attr-Sim): These methods, rooted in Case-Based Reasoning (CBR), use "Nearest Unlike Neighbors" (NUNs) to guide the search for semi-factuals.
- Local-Region Method: This approach uses a local surrogate model (like logistic regression) to capture the query's neighborhood and identify marginal instances near the decision boundary.
- Most Distant Neighbors (MDN) Method (Novel): This new method proposes finding the most distant same-class instance from the query on a key-feature. It uses a "Semi-Factual Scoring" (sfs) function that prioritizes instances with sparse feature differences and significant value differences on a key-feature.
The benchmark study involved leave-one-out cross-validation across seven tabular datasets, evaluating methods on six metrics: Query-to-SF Distance, Query-to-SF KNN (%), SF-to-Query-Class Distance, SF-to-NUN Distance, MDN Distance, and Sparsity.
Datasets
The benchmark study utilized seven standard, publicly available, binary-classed tabular datasets commonly used in counterfactual literature:
- Adult-Income
- Blood Alcohol
- Default Credit Card
- Pima Diabetes
- German Credit
- HELOC
- Lending Club
The paper does not introduce new datasets but rather uses existing ones to benchmark the proposed and historical methods.
Key Insights and Ideas
- Psychological Distinction: Semi-factuals have demonstrably different psychological impacts than counterfactuals. While counterfactuals encourage users to consider alternative actions that undo an outcome, semi-factuals reinforce the status quo and weaken the perceived causal link between features and outcomes. This makes them valuable for explaining why a decision is robust or inevitable.
- Desiderata for Semi-Factual XAI: The paper articulates a comprehensive set of desiderata, including factual correctness, sparse feature changes, plausibility (within the data manifold), convincingness (even if surprising), impact on user perception of causality, and fairness/ethical considerations (avoiding proxies, ensuring robustness). This provides a much-needed framework for evaluating semi-factual explanation methods.
- MDN Method's Promise: The novel MDN method shows strong performance on metrics related to distance from the query and finding "good" semi-factuals (as measured by the sfs function), suggesting it's effective at identifying instances that are significantly different from the query within the same class.
- Trade-offs in Sparsity: While methods like MDN and Local-Region are good at finding distant semi-factuals, they tend to produce explanations with more feature differences (lower sparsity). This highlights a potential trade-off between explanation distance and simplicity.
Key Results
- MDN and Local-Region Lead: The MDN and Local-Region methods generally performed best across the benchmarked metrics, particularly on measures of distance from the query and identifying good semi-factuals. MDN excelled on Query-to-SF Distance, Query-to-SF KNN, and MDN Distance.
- KLEOR Variants' Niche: The KLEOR variants (Sim-Miss, Global-Sim, Attr-Sim) performed relatively better on the SF-to-NUN Distance metric, as they are optimized to find instances close to the class boundary.
- Sparsity Challenges: The MDN method, in particular, showed lower sparsity (more feature differences) on several datasets, indicating that while it finds distant semi-factuals, these might be less concise. The authors suggest fine-tuning the sfs function could improve this.
- Historical Methods' Limitations: The paper notes that historical methods like Doyle et al.'s utility-based approach are knowledge-intensive and less suitable as general benchmarks.
Limitations and Future Work
- Tabular Data Focus: The current benchmark study is limited to tabular data, although the authors believe the desiderata and metrics are applicable to image and time-series data.
- Paucity of User Studies: A significant limitation is the lack of extensive user studies. The authors emphasize that understanding how users actually respond to and interpret semi-factual explanations is crucial for their practical adoption.
- Ethical Considerations: The paper reiterates the ethical implications of semi-factuals, noting their cognitive impact and potential for misuse if assumptions (like robustness) are violated. Future work should focus on developing metrics to audit these assumptions.
- Improving MDN Sparsity: The authors suggest that fine-tuning the sfs function in the MDN method could improve its sparsity performance.
- Broader Applicability: Future research should explore applying these methods and metrics to other data modalities and conducting rigorous user studies to validate their effectiveness and impact.
Summary updated