XALIGN: A Human-centric Visual Analytics Approach to Understanding AI Explanations

The hybrid rationale is well argued and the matrix visualization effectively reveals tensions that single-method views obscure. In the German Credit study, the interface exposes a direct contradiction on \u201cOther installment plans\u201d where all three attribution methods conflict with the rule-based direction, and the meteorology case shows that high rule counts can coexist with weak attributions because a feature may split many paths without shifting prediction probability. As the authors note, \u201cThe rule count shows how much the model depends on the feature to partition the classes whereas the attributions show how much the feature moves the prediction towards or against a particular class.\u201d (Kathirgamanathan et al., Section 4.3).

The evaluation is too small to support the paper's claims. Only three experts were recruited, there was no comparison interface, and the cross-class comparison tab was added after the study and never evaluated; the authors concede that \u201cLimitations of the study include the small number of participants and no other comparison to XALIGN.\u201d (Kathirgamanathan et al., Section 5.3). A second, deeper issue is the reliance on purely subjective pattern matching: \u201cAgreement is judged visually. We do not compute a quantitative agreement measure.\u201d (Kathirgamanathan et al., Section 3.1, Step 5). This makes the approach hard to scale and does not guarantee that different users would perceive the same agreements.\n\nThe scope is also narrower than the title implies. XALIGN is tightly bound to Random Forests because its rule extraction traverses tree paths and TreeSHAP is model-specific; the authors acknowledge that RF is \u201cAccurate but hard to interpret directly; its tree structure allows rule extraction and TreeSHAP.\u201d (Kathirgamanathan et al., Fig. 1). Positioning the work as a general approach to \u201cUnderstanding AI Explanations\u201d overreaches when the entire pipeline is built for a single model class.

The related work discussion is careful and fair, correctly separating XALIGN from attribution-only tools such as Visagreement and ExplainExplore and from model-inspection dashboards like Manifold and the What-If Tool. However, the paper never tests against a simpler baseline\u2014e.g., a SHAP summary plot paired with a decision tree\u2014to demonstrate that the coordinated matrix actually improves interpretation. The case studies remain illustrative: the authors assert that \u201cagreement serves as a practical indicator of explanation robustness\u201d (Kathirgamanathan et al., Section 4.3), but without ground-truth importance or causal verification this is an assumption, not an established fact.

Reproducibility is weak. No code repository or software release is mentioned. While the German Credit dataset is public and the meteorological data are identified by DOI, critical explainer hyperparameters\u2014such as the number of background samples for KernelSHAP, perturbation counts for LIME, or kernel widths\u2014are omitted entirely. The rule extraction and aggregation pipeline is described only at the level of \u201cTraverse every tree (each leaf gives a rule), then clean, prune and aggregate the rules.\u201d (Kathirgamanathan et al., Fig. 1), without pseudocode or implementation detail, so an independent replication would require substantial reverse engineering.