Beyond Feature Importance: A Comparative Analysis of Pattern Detection Methods in Cluster Interpretation

Published: Accepted for 2026 36th Irish Signals and Systems Conference (ISSC), arXiv:2608.05880v1 [cs.LG]

---

The paper addresses a fundamental challenge in cluster analysis: interpreting why a cluster exists and identifying the structured patterns that define it. The authors note that while clustering algorithms can successfully partition data into groups, the analytical value of clustering lies not only in assigning data points to groups but in identifying meaningful patterns that characterise those groups. This is particularly important in high-dimensional domains such as biomedical research and population health analysis.

A key obstacle identified is the absence of ground truth in real datasets: Without known patterns, it is difficult to assess whether an explanation method successfully captures meaningful structure or merely produces plausible interpretations. The authors argue that existing interpretability techniques—feature importance from surrogate models, LIME, and PCA—"were not originally designed for pattern detection within clustering outcomes. Instead, they focus on identifying influential variables, explaining individual predictions, or capturing variance in reduced representations."

The paper formally defines the pattern detection problem as distinct from clustering and feature attribution. A pattern P is defined as a subset of features and a joint feature behaviour satisfying one or more of:

1. Value-based patterns: Features in FP take consistently high or low values across a majority of samples.

2. Correlation-based patterns: Features in FP are statistically dependent, either positively or negatively, within the cluster: Corr(Xi, Xj) ≠ 0, ∀i, j ∈ FP.

3. Distributional patterns: Features in FP exhibit heterogeneous distributions, e.g., multimodal behaviour.

The goal of a pattern detection method is to output a set of patterns identifying the subset of features FPi involved, the type of joint behaviour (value, correlation, distributional), and the strength or prominence of the pattern within the cluster.

To enable controlled evaluation, the authors constructed a suite of synthetic datasets. Patterns were defined in collaboration with medical domain experts to reflect behaviours commonly observed in clinical datasets. Each dataset represents a single cluster from an upstream clustering algorithm, thereby isolating the task of pattern detection from clustering itself.

Six pattern types were considered:

• High Values: features near upper end of distribution (2-3 features)

• Low Values: features near lower end of distribution (2-3 features)

• Positive Correlation: positive linear correlation (2-4 features)

• Negative Correlation: negative linear correlation (2-3 features)

• High-Low: combination of 2 high and 2 low features

• Switch High–Low: approximately half of the samples take high values and the remaining half take low values (2 or 4 features)

Each dataset has n=500 samples and dimensionality d ∈ 10, 20, 30, 40, 50, 60, 70, 80, 90, 100. All features were initially independently generated from a normal distribution scaled to [0,1], ensuring any detectable patterns arise solely from the controlled injection procedure. For each selected feature, 80% of data points were modified to satisfy pattern constraints, leaving 20% unchanged to introduce controlled noise. A total of twelve unique pattern combinations were constructed across all dimensionalities, yielding 120 datasets.

Three methods were evaluated:

RF Surrogate Model with Permutation Feature Importance: Cluster labels are treated as targets, and a Random Forest classifier is trained in a one-versus-rest formulation. Permutation importance measures the decrease in predictive performance when feature values are randomly permuted. The authors note that "permutation importance captures global predictive relevance and supports nonlinear relationships learned by the model. However, feature importance identifies influential variables but does not directly reveal structured multi-feature patterns."

LIME: A local explanation framework that approximates a complex model with an interpretable model within the neighbourhood of a specific instance. The procedure involves generating synthetic samples, querying the black-box model, weighting by proximity, and fitting an interpretable model with discretized features. The authors note LIME provides "interpretable local explanations and model-agnostic applicability. However, explanations are instance-specific and sensitive to sampling variability, making aggregation into cluster-level pattern descriptions challenging."

PCA: A linear dimensionality reduction technique that identifies orthogonal directions maximising variance through eigenvalue decomposition of the covariance matrix. High-magnitude loadings indicate importance; similar signs indicate positive correlation, while opposite signs indicate negative correlation. The authors note PCA assumes linear relationships and prioritises variance maximisation rather than semantic pattern identification.

Both methods assign importance scores to features, evaluated using top-m accuracy (ratio of pattern features ranked in the top positions to total pattern features) and a margin metric (difference between min importance of pattern features and max importance of non-pattern features).

RF Surrogate Model Findings:

• Achieved consistently high ranking accuracy across most pattern types, typically exceeding 0.95 and often reaching perfect identification of top pattern features.

• Ranking accuracy remains high across all dimensionalities, indicating robustness to increasing feature-space complexity.

• Produced small but positive margins, indicating that relevant features are usually ranked above irrelevant ones, albeit with limited separation—meaning there is considerable overlap between importance scores with no clear threshold separating pattern members from irrelevant variables.

PCA Findings:

• Achieved perfect or near-perfect ranking accuracy for several patterns, particularly for positive, negative, and switching structures.

• Exhibited substantially more variability, with performance dropping markedly for certain configurations, most notably the Low patterns (ranking accuracy as low as 0.05).

• Produced large positive margins for some patterns (e.g., Positive and Negative patterns), suggesting strong separation, but frequently yields negative margins, particularly for Low and mixed High–Low patterns, indicating non-pattern features occasionally ranked higher than true pattern features.

• Showed a clear decline in ranking accuracy as data dimensionality increases, attributed to variance dilution, where informative signals are overshadowed by accumulated variability from irrelevant dimensions.

Critical shared limitation: "Neither explicitly reveals the relationships among detected features... both RF and PCA reduce the output to individual feature importance scores. As a result, the methods identify which features appear important but do not clarify whether these features belong to a coherent pattern or simply exhibit independent statistical influence. This is particularly evident for higher-order patterns, where multiple interacting features define the signal but no explicit grouping or relational structure is recovered. Consequently, successful ranking does not necessarily correspond to successful pattern discovery."

LIME was evaluated by examining its four highest-ranked rules and associated weights for each pattern type. Findings include:

• LIME consistently places key pattern features within the top explanatory rules, indicating that the method successfully identifies variables that influence the predictions.

• No clearly irrelevant features appear among the explanations; however, individual rules typically capture only subsets of the pattern features rather than representing the full structure jointly.

• The presence of both positive and negative rules for the same feature does not imply inconsistency in the underlying model but rather reflects locally opposing contributions across neighbouring value regions. Nevertheless, this behaviour complicates interpretation.

• Patterns defined by simple magnitude effects can be partially inferred from the rules, since threshold conditions with strong positive weights often align with the injected feature behaviour.

• Correlation-based patterns cannot be directly extracted because LIME expresses explanations as independent univariate conditions and does not encode relationships between features.

• Switching patterns are only weakly represented: while individual high or low value conditions are detected, the coordinated opposing behaviour among feature groups is not captured as a unified rule.

• Overall, pattern information becomes fragmented across multiple locally valid explanations, requiring substantial manual interpretation to reconstruct the intended structure.

The authors conclude that while all three approaches can highlight important variables to some extent, none of them reliably identifies complete multi-feature patterns or clearly captures relationships among features.

Key conclusions:

1. The RF surrogate model showed strong and consistent ranking performance across pattern types and data dimensionalities, but its outputs were limited to feature importance rankings, providing little insight into how features relate to one another or collectively form patterns.

2. PCA exhibited strong performance when patterns aligned with dominant variance directions but degraded substantially as dimensionality increased or when patterns did not dominate global variance, proving sensitive to the statistical structure of the data and unreliable for detecting more complex or subtle patterns.

3. LIME produced fragmented valid rules that did not coherently represent global feature relationships, with correlation-based and switching patterns could not be directly recovered due to the method's reliance on independent univariate conditions.

The broader methodological implication is that current widely used techniques are primarily designed for feature ranking, dimensionality reduction, or local explanation rather than true pattern detection. This "has important implications for downstream tasks such as cluster analysis and exploratory data analysis, where meaningful grouping often depends on detecting interactions and coordinated behaviour among features rather than assessing variables independently."

The paper's contributions are threefold:

1. A formal framing of pattern detection as a distinct task from clustering and feature attribution.

2. A controlled synthetic benchmark for evaluating pattern detection methods with known ground truth.

3. A comparative analysis demonstrating systematic limitations of widely used interpretability techniques, motivating the need for dedicated pattern detection methodologies.

The authors state: "Future work should focus on developing new approaches capable of explicitly identifying multi-feature patterns and modelling relationships among variables. Such methods should move beyond independent feature scoring toward representations that capture collective behaviour, structural dependencies, and interpretable pattern boundaries. They also note that evaluation on real-world datasets was beyond the scope of this work, and future work should expand evaluation to real-world datasets, which would also provide a comparative baseline for future pattern detection methods to be evaluated against, further improving confidence in novel approaches."

Specifically, AI systems can be improved in the following ways based on this paper's findings:

• Explicit multi-feature pattern extraction: Instead of outputting per-feature importance scores, the improved AI system will detect and output complete patterns—subsets of features with a specified joint behavior type (e.g., high-value, low-value, positive/negative correlation, switching, or distributional), along with the strength and reliability of each pattern. This directly addresses the paper's central limitation that current methods reduce the output to individual feature importance scores and fail to reveal coherent multi-feature structures.

• Joint relationship modeling rather than independent scoring: The improved system will explicitly model pairwise or higher-order interactions among features within each cluster, using covariance/correlation structure, graphical models, or copula-based dependency measures. It will be able to distinguish between features that are independently statistically significant and features that jointly form a pattern, tackling the identified failure where successful ranking does not necessarily correspond to successful pattern discovery.

• Correlation and switching pattern detection: The system will go beyond univariate threshold rules and directly learn coordinated behavior among features. For example, it will detect that two or more features are positively or negatively correlated within a cluster, or that a group of features switches between high and low states across samples, rather than fragmenting these into local independent rules as LIME does.

• Robustness to high-dimensional variance dilution: The improved system will incorporate dimensionality-aware regularization, sparse latent variable models, or targeted subspace search so that pattern detection remains reliable as the number of irrelevant features grows. This counters the observed PCA degradation where informative signals are overshadowed by accumulated variability from irrelevant dimensions.

• Global aggregation with local explanation consistency: When using local explainers like LIME, the improved system will aggregate many local explanations into a cluster-level representation that preserves relational structure, resolves contradictory local rules, and avoids fragmentation. It will produce a unified pattern description rather than requiring substantial manual interpretation.

• Validation against known ground-truth patterns: The improved system will be trained and evaluated on the paper's synthetic benchmark (or similar controlled datasets with known patterns), allowing it to measure not just top- m ranking accuracy but also the correctness of the recovered joint structure, pattern boundaries, and relationship types—thereby providing guarantees about pattern detection quality.

• Uncertainty and confidence metrics for discovered patterns: The system will output confidence scores or statistical significance for each detected pattern, enabling users to distinguish reliable structural findings from plausible-but-weak artifacts, especially in the absence of ground truth in real datasets.

• End-to-end cluster interpretation pipeline: The improved system will integrate clustering with pattern detection, taking raw cluster assignments and automatically producing an interpretable summary of each cluster's defining patterns, including feature subsets, behavior types, and visualizable relationship graphs—directly supporting downstream exploratory data analysis and domain decision-making.