Poster for "DualEx: Dual-Space Clustering for Regional Explanations"

DualEx: Dual-Space Clustering For Regional Explanations Regional Explanations and Model Bias Method Model bias occurs when different models produce different regional explanations on the same data. This variation, seen in NMI and ARI, reflects model bias, where the explanation space is driven by model behavior rather than data structure. Figure 1 Feature importance levels: local describe one patient, global summarise all, and regional reveal regions with consistent feature patterns. . Regional explanations bridge the gap between local and global explanations by identifying coherent groups of instances that share similar model behavior. Each region summarises both what the model focuses on and why, providing insight at a meaningful intermediate level. ➔Given xi and xj , we define a distance metric balancing data space robustness and explanation space clustering quality. Figure 2 The data space is independent of the ML model providing thus maximum robustness. Instances are generally noisier than explanation vectors. Pernille Matthews, Tommaso Amico, Arthur Zimek, Ira Assent Figure 3 The dual space retains the data space robustness while inheriting the explanationʼs space separability and compactness. Figure 4 The explanation space, by leveraging explanation vectorsʼ information, finds separable, compact and domain aligned regions. Results Establishing regional explanations solely in the explanation space induces model bias. In the data space, meaningful separation is difficult. We introduce DualEx, a method that bridges these extremes by combining both spaces for robust and interpretable regional explanations. Table 1 Model agreement across explanation spaces. Low NMI and ARI reveal how different models produce inconsistent regional explanations; evidence of model bias. Results across 40 OpenML datasets, quantifying robustness as the average pairwise NMI, shows that explanations are robust at low alphas. As the explanation spaceʼs contribution rises explanations become inconsistent across models. Case study showing the Silhouette Score (clustering quality) rises with 𝛂, due to natural compactness of the explanation space, and average NMI (robustness) decreases with alpha, due to model-bias. Figure 5 Trade-off between clustering quality and robustness. Higher 𝛂 improves regional cohesion but adds model-specific bias. DualEx balances data and explanation spaces while still obtaining meaningful regional explanations. Table 2 Average Pairwise NMI across three cluster evaluation metrics SC, DBCV, ARI, illustrating how model-bias leads to a decrease in coherence and robustness across models. ➔Clustering algorithms are applied to derive coherent and high-quality regions reflecting both data structure and model reasoning. ➔ α regulates the contribution of each space, transitioning from data-driven to explanation-driven clustering outcomes.