Paper deep dive
Beyond Feature Importance: A Comparative Analysis of Pattern Detection Methods in Cluster Interpretation
Benjamin Connor, Anna Jurek-Loughrey, Lu Bai, Muhammad Fahim
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/8/2026, 4:18:58 AM
Summary
This paper evaluates the effectiveness of three post-hoc explainability methods—Random Forest (RF) surrogate models, LIME, and Principal Component Analysis (PCA)—for detecting structured patterns within clustering results. Using synthetic datasets with injected ground-truth patterns (value-based, correlation-based, and distributional), the study finds that while these methods can identify relevant features, they fail to consistently detect multi-feature patterns or relationships between features. The work highlights a gap between existing explainability tools and the requirements for pattern-level cluster interpretation, motivating the development of dedicated pattern detection methodologies.
Entities (8)
Relation Signals (9)
Random Forest → evaluatedfor → Pattern Detection
confidence 95% · Three widely used techniques are evaluated: a Random Forest surrogate model... for pattern detection in clustering results.
LIME → evaluatedfor → Pattern Detection
confidence 95% · Three widely used techniques are evaluated: ... LIME ... for pattern detection in clustering results.
Principal Component Analysis → evaluatedfor → Pattern Detection
confidence 95% · Three widely used techniques are evaluated: ... and principal component analysis.
Random Forest → uses → Permutation Feature Importance
confidence 95% · A Random Forest surrogate model with permutation feature importance...
Clustering → challenges → Pattern Detection
confidence 90% · Interpreting clustering outcomes remains a fundamental challenge... particularly in domains such as healthcare where meaningful patterns must be extracted...
Synthetic Datasets → usedby → Pattern Detection
confidence 90% · To enable controlled evaluation, we introduce a suite of synthetic datasets in which predefined patterns are systematically injected.
Random Forest → limitations → Pattern Detection
confidence 85% · Results demonstrate that although each method can successfully recover relevant features, none consistently detects all injected pattern types.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Interpreting clustering outcomes remains a fundamental challenge in data analysis, particularly in domains such as healthcare where meaningful patterns must be extracted from high-dimensional data. While numerous explainability techniques exist, they are primarily designed to assess feature importance or provide local instance-level explanations rather than to identify structured patterns present within clusters. This work presents a comparative evaluation of commonly used post-hoc analysis methods for pattern detection in clustering results. To enable controlled evaluation, we introduce a suite of synthetic datasets in which predefined patterns are systematically injected. Three widely used techniques are evaluated: a Random Forest surrogate model with permutation feature importance, LIME (Local Interpretable Model-agnostic Explanations), and principal component analysis. Results demonstrate that although each method can successfully recover relevant features, none consistently detects all injected pattern types. These findings high- light a critical gap between existing explainability tools and the requirements of pattern-level cluster interpretation, motivating the development of dedicated pattern detection methodologies.
Tags
Links
- Source: https://arxiv.org/abs/2608.05880v1
- Canonical: https://arxiv.org/abs/2608.05880v1
Trouble viewing inline? Open PDF directly →
Full Text
33,084 characters extracted from source content.
Expand or collapse full text
Beyond Feature Importance: A Comparative Analysis of Pattern Detection Methods in Cluster Interpretation Benjamin Connor Anna Jurek-Loughrey Lu Bai Muhammad Fahim Abstract Interpreting clustering outcomes remains a fundamental challenge in data analysis, particularly in domains such as healthcare where meaningful patterns must be extracted from high-dimensional data. While numerous explainability techniques exist, they are primarily designed to assess feature importance or provide local instance-level explanations rather than to identify structured patterns present within clusters. This work presents a comparative evaluation of commonly used post-hoc analysis methods for pattern detection in clustering results. To enable controlled evaluation, we introduce a suite of synthetic datasets in which predefined patterns are systematically injected. Three widely used techniques are evaluated: a Random Forest surrogate model with permutation feature importance, LIME (Local Interpretable Model-agnostic Explanations), and principal component analysis. Results demonstrate that although each method can successfully recover relevant features, none consistently detects all injected pattern types. These findings highlight a critical gap between existing explainability tools and the requirements of pattern-level cluster interpretation, motivating the development of dedicated pattern detection methodologies. © 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. This work has been accepted for publication in 2026 36th Irish Signals and Systems Conference (ISSC). The final published version will be available via IEEE Xplore. I Introduction Clustering algorithms are widely used to uncover latent structure in complex datasets, particularly in high-dimensional domains such as biomedical research, population health analysis, and exploratory data mining. While clustering methods can successfully partition data into groups with similar characteristics, interpreting why a cluster exists and what patterns define it remains a major unresolved challenge [10, 2]. In practice, the analytical value of clustering lies not only in assigning data points to groups but in identifying meaningful patterns that characterise those groups. Existing approaches to cluster interpretation typically rely on post-hoc explainability techniques. Feature importance measures derived from surrogate models, local explanation frameworks such as LIME, and dimensionality reduction techniques such as principal component analysis (PCA) are frequently applied to understand cluster structure [7, 9, 8, 14]. However, these methods 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. As a result, their ability to recover structured multi-feature patterns remains unclear. In many real-world applications, particularly in medical research, clusters are expected to represent interpretable behavioural or physiological patterns rather than isolated feature effects [13, 7]. Such patterns may involve coordinated feature behaviour, including correlated variables, jointly elevated or suppressed values, or heterogeneous distributions within a cluster. Detecting these structures requires methods capable of identifying relationships among features rather than evaluating them independently. A key obstacle in evaluating pattern detection methods 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. To address this limitation, we construct a controlled experimental framework based on synthetic datasets with explicitly injected patterns. Each dataset represents a single cluster generated from normally distributed features, into which predefined patterns are introduced while preserving realistic noise levels. This design enables systematic comparison between detected explanations and known underlying structure. In this study, we evaluate three commonly used approaches for analysing clustering outcomes: Random Forest (RF) surrogate model using permutation feature importance, LIME and PCA. Each method is applied independently to synthetic clusters containing known pattern types, including high-value, low-value, correlation-based, and distribution-switching patterns. Through this evaluation, we examine not only whether relevant features are identified but whether relationships between features-and therefore patterns-are successfully recovered. Our findings reveal that while these approaches highlight important variables, they fall short of reliably detecting multi-feature patterns, underscoring the need for dedicated pattern detection methods. The contributions of this work 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. By establishing these limitations empirically, this work lays the foundation for future research aimed at developing methods specifically designed to detect and quantify patterns within clustering results. I Background and Related Work Clustering algorithms aim to discover latent structure in data by grouping observations according to similarity. While substantial research has focused on improving clustering algorithms themselves [5], comparatively less attention has been devoted to interpreting the resulting clusters. In many applied domains, particularly healthcare and biomedical analytics, the primary objective is not the cluster assignment itself but the identification of patterns that characterise each cluster. Cluster interpretation is commonly performed using summary statistics, centroid analysis, or feature ranking approaches [12, 4, 7]. These techniques describe differences between clusters but often treat features independently, limiting their ability to reveal structured relationships such as correlated feature groups or heterogeneous distributions within a cluster. Consequently, researchers increasingly employ model explainability methods originally developed for supervised learning to analyse clustering outcomes. A common strategy is to transform clustering into a supervised problem by training a classifier to predict cluster membership and then applying explainability tools to the classifier [1]. While practical, this approach implicitly assumes that methods designed for prediction explanation are suitable for pattern discovery - an assumption that has not been systematically validated. I-A Surrogate Models and Feature Importance Surrogate models approximate the behaviour of a complex or non-interpretable system using a more interpretable predictive model [3]. In the context of clustering interpretation, cluster labels are treated as targets, and a supervised model is trained to predict (one-versus-all) cluster membership. Explanations are then derived from the surrogate model. Let the dataset be defined as X∈ℝn×dX ^n× d, where n denotes the number of samples and d the number of features, with cluster labels y∈0,1y∈\0,1\. A surrogate model learns a mapping f:X→yf:X→ y that approximates the cluster decision boundary. In this work, a RF classifier is employed in a one-versus-rest formulation. Feature importance is computed using permutation importance, which measures the decrease in predictive performance when feature values are randomly permuted. 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. I-B Local Interpretable Model-Agnostic Explanations (LIME) LIME is a local explanation framework designed to explain individual predictions of black-box models using locally faithful interpretable approximations [11]. Given an instance x, LIME approximates a complex model f with a simpler interpretable model g within the neighbourhood of x. The procedure consists of: 1. Choosing a specific data point x we want to explain, 2. Generating synthetic samples S around instance x, 3. Querying the black-box model for predictions on x and the samples from S, 4. Weighting samples in S according to proximity to x, 5. Fitting an interpretable model on S, 6. Extracting feature contributions as explanation rules. Continuous variables are typically discretized into bins to produce human-readable rules describing feature ranges. In our evaluation, discretization is applied to all continuous features, with explanations restricted to a fixed number of features and generated using a predefined number of synthetic samples. 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. I-C Principal Component Analysis (PCA) PCA is a linear dimensionality reduction technique widely used for exploratory data analysis and structure discovery [6]. Given a centred data matrix X, PCA identifies orthogonal directions that maximise variance: max‖w‖=1Var(Xw). _\|w\|=1Var(Xw). This optimisation leads to the eigenvalue decomposition of the covariance matrix Σ=1nXTX. = 1nX^TX. Eigenvectors of Σ define principal components, while eigenvalues represent explained variance. Each principal component is computed as zk=Xwk,z_k=Xw_k, where wkw_k denotes the loading vector associated with component k, which indicate contribution strength. High-magnitude loadings indicate higher importance. Similar signs indicate positive correlation, while opposite signs indicate negative correlation. PCA effectively captures linear correlation structures among features. However, it assumes linear relationships and prioritises variance maximisation rather than semantic pattern identification. It does not guarantee that the resulting components correspond to meaningful or interpretable patterns. I Problem Definition Clustering algorithms partition a dataset X∈ℝn×dX ^n× d into K clusters =C1,C2,…,CK,⋃k=1KCk=X,C=\C_1,C_2,…,C_K\, _k=1^KC_k=X, such that points within a cluster are more similar to each other than to points in other clusters [5]. While clustering assigns membership, it does not explain why clusters exist or which feature combinations define them [2]. We define the pattern detection problem as the task of identifying structured, multi-feature relationships within a given cluster that distinguish it from other clusters or from random noise. I-A Definition of a Pattern Let CkC_k denote a cluster and XCk∈ℝ|Ck|×dX_C_k ^|C_k|× d its associated data. A pattern P is defined as a subset of features FP⊆1,2,…,dF_P \1,2,…,d\ and a joint feature behaviour over XCkX_C_k that satisfies one or more of the following properties: 1. Value-based patterns: Features in FPF_P take consistently high or low values across a majority of samples. 2. Correlation-based patterns: Features in FPF_P are statistically dependent, either positively or negatively, within the cluster: Corr(Xi,Xj)≠0,∀i,j∈FP.Corr(X_i,X_j)≠ 0, ∀ i,j∈ F_P. 3. Distributional patterns: Features in FPF_P exhibit heterogeneous distributions, e.g., multimodal behaviour. I-B Objectives of Pattern Detection Given cluster CkC_k, the goal of a pattern detection method is to output a set of patterns k=P1,P2,…,PNk,P_k=\P_1,P_2,…,P_N_k\, where each pattern PiP_i identifies: • The subset of features FPiF_P_i involved, • The type of joint behaviour (value, correlation, distributional), • The strength or prominence of the pattern within the cluster. I-C Evaluation Context The problem is evaluated under a controlled synthetic framework where clusters contain known patterns of varying types (value-based, correlation-based, switch/multimodal). This allows rigorous assessment of whether existing methods (surrogate models, LIME, PCA) can correctly identify patterns discussed above. IV Synthetic Data Design In order to verify that pattern detection methods reliably identify meaningful structures within clustering results, a suite of controlled synthetic datasets was constructed. Evaluation on real-world datasets alone is insufficient for this purpose, as the true underlying patterns are typically unknown. Synthetic data enables the introduction of predefined ground-truth patterns, allowing objective assessment of whether an analysis method successfully recovers known structure. Patterns were defined in collaboration with medical domain experts to reflect behaviours commonly observed in clinical datasets, including coordinated feature behaviour, extreme value characteristics, and heterogeneous distributions within patient subgroups. Each synthetic dataset was designed to represent the outcome of a single cluster produced by an upstream clustering algorithm, thereby isolating the task of pattern detection from clustering itself. Five categories of patterns were considered: • High Values: Selected features take values near the upper end of their distribution (2-3 features). • Low Values: Selected features take values near the lower end of their distribution (2-3 features). • Positive Correlation: Selected features exhibit positive linear correlation (2-4 features). • Negative Correlation: Selected features exhibit negative linear correlation (2-3 features). • High-Low Some features take values either near the upper or lower end of their distribution (2 high, 2 low). • Switch High–Low: Selected features display heterogeneous behaviour, where approximately half of the samples take high values and the remaining half take low values (2 or 4 features). Each pattern involves a predefined number of randomly selected features, ranging from two to four, as shown in the brackets. Each synthetic dataset consists of n=500n=500 samples and a varying number of features d∈10,20,30,40,50,60,70,80,90,100d∈\10,20,30,40,50,60,70,80,90,100\. Prior to pattern injection, all features were independently generated from a normal distribution scaled to the interval [0,1][0,1]. This initialisation produces datasets without intrinsic structure, ensuring that any detectable patterns arise solely from the controlled injection procedure. For each dataset, a pattern type was selected from the predefined list and injected to randomly selected features. For each selected feature, 80%80\% of data points were modified to satisfy the pattern constraints and the remaining 20%20\% of samples were left unchanged to introduce controlled noise. This design simulates realistic clustering outcomes in which patterns are prominent but not perfectly expressed across all observations. A total of twelve unique pattern combinations were constructed. Each combination was generated across all feature dimensionalities, resulting in 120120 datasets. V Experimental Results In this work we evaluate three pattern detection methods discussed earlier in the paper. These include RF surrogate model and the PCA method, which both assign importance scores to features, and the LIME method, which outputs a collection of weighted rules. For RF surrogate model and the LIME method, the predictive task was defined as one-vs-all classification, where each of the patterns is considered as an individual cluster. The results obtained by each of the three methods on our synthetic dataset are explained in the following sections. V-A RF Surrogate Model and PCA Tables I and I summarise the results obtained by the RF surrogate models and the PCA method across all datasets. Both approaches assign importance scores to features, which are ranked to evaluate whether features participating in the true patterns are correctly prioritised. The first column lists the type of the pattern. For each pattern, the results are averaged across the different dataset dimensions. The second column reports the evaluation of the feature ranking using top−mtop-m accuracyaccuracy score, calculated as a ratio between the number of pattern features put at the top of the ranking and the total number of pattern features. This is followed by the standard deviation in column three. The fourth column shows the average difference between the min importance values assigned to a pattern feature and the max importance value assigned to a non-pattern feature. The final column provides the standard deviation of the values reported in column 4. TABLE I: Summary of the results obtained for the RF surrogate model. Pattern Ranking Accuracy St. Dev. Margin St. Dev. Positive 2 0.95 0.1581 0.0601 0.0488 Negative 2 1.0 0.0 0.1193 0.0465 High 2 1.0 0.0 0.3048 0.035 Low 2 0.95 0.1581 0.138 0.1137 Switch 2 1.0 0.0 0.2253 0.087 Positive 3 0.9667 0.1054 0.0191 0.0288 Negative 3 0.9667 0.1054 0.0181 0.0589 High 3 0.9667 0.1054 0.0291 0.0391 Low 3 1.0 0.0 0.0131 0.0126 Positive 4 0.925 0.1687 0.0054 0.0117 High-Low 4 0.875 0.1318 0.0012 0.0061 Switch 4 0.875 0.1768 0.0144 0.0381 TABLE I: Summary of the results obtained for the PCA method. Pattern Ranking Accuracy St. Dev. Margin St. Dev. Positive 2 1.0 0.0 0.4424 0.0815 Negative 2 1.0 0.0 0.4477 0.0681 High 2 0.8 0.4216 0.089 0.2279 Low 2 0.05 0.1581 -0.2742 0.123 Switch 2 0.8 0.4216 0.2106 0.2304 Positive 3 1.0 0.0 0.4459 0.0345 Negative 3 0.7 0.1892 -0.1238 0.0815 High 3 1.0 0.0 0.3373 0.0542 Low 3 0.2667 0.3443 -0.223 0.07 Positive 4 1.0 0.0 0.3971 0.0273 High-Low 4 0.525 0.1845 -0.1504 0.0742 Switch 4 1.0 0.0 0.3277 0.0488 The RF-based method achieves consistently high ranking accuracy across most pattern types, typically exceeding 0.95 and often reaching perfect identification of top pattern features. This indicates that RF reliably assigns high importance to features associated with injected signals. PCA also achieves perfect or near-perfect ranking accuracy for several patterns, particularly for positive, negative, and switching structures. However, PCA exhibits substantially more variability, with performance dropping markedly for certain configurations, most notably the Low patterns. This contrast suggests that RF provides more stable feature prioritisation across pattern types, whereas PCA is sensitive to the statistical structure of the signal and performs well primarily when patterns align with dominant variance directions. A key difference between the two methods emerges when examining the margin metric, which measures the separation between pattern and non-pattern features. RF generally produces small but positive margins, indicating that relevant features are usually ranked above irrelevant ones, albeit with limited separation. The small margins imply considerable overlap between importance scores, meaning that there is no clear threshold distinguishing pattern members from the remaining variables. In contrast, PCA produces large positive margins for some patterns (e.g., Positive and Negative patterns), suggesting strong separation when the underlying structure aligns with principal variance components. However, PCA also frequently yields negative margins, particularly for Low and mixed High–Low patterns. Negative margins indicate that non-pattern features are occasionally ranked higher than true pattern features, demonstrating instability and reduced robustness across pattern types. Thus, RF shows consistent but weak separation, while PCA exhibits stronger separation when successful but fails more dramatically when assumptions are violated. Despite relatively high ranking accuracy, both methods share an important limitation: neither explicitly reveals the relationships among detected features. The injected patterns encode structured behaviour - correlations, shared magnitude shifts, or coordinated switching dynamics - yet 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 limitation 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. We also investigate how the two methods perform across datasets of varying dimensionality. Figures 1 and 2 present the ranking accuracy as the number of features increases. The results show that the RF surrogate model maintains consistently high performance across all examined dimensionalities, indicating robustness to increasing feature-space complexity. This suggests that the RF-based importance estimation effectively isolates informative variables even when the number of irrelevant features grows. In contrast, PCA exhibits a clear decline in ranking accuracy as data dimensionality increases. This degradation can be attributed to PCA’s reliance on variance maximisation: as additional noisy or unrelated features are introduced, the dominant variance directions become increasingly influenced by non-pattern features. Consequently, the principal components capture global variance structure rather than the injected patterns, leading to reduced ability to prioritise relevant variables. These results highlight a fundamental difference between the two approaches: RF surrogate models evaluate feature importance through predictive relationships that remain relatively stable in high-dimensional settings, whereas PCA becomes progressively sensitive to variance dilution, where informative signals are overshadowed by accumulated variability from irrelevant dimensions. Figure 1: Chart showing mean ranking accuracy of RF surrogate model across synthetic datasets with increasing number of dimensions. Figure 2: Chart showing mean ranking accuracy of PCA method across synthetic datasets with increasing number of dimensions. V-B LIME Table I summarises the explanations produced by the LIME method for datasets containing injected feature patterns, reporting the four highest-ranked rules and their associated weights for each case. Across all patterns, LIME consistently places key pattern features within the top explanatory rules, indicating that the method successfully identifies variables that influence the predictions. Importantly, 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 interpretation of rule weights is therefore critical: positive weights indicate feature value ranges that support the prediction, whereas negative weights correspond to ranges that oppose it within the surrogate linear model fitted by LIME. Consequently, 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, as some explanations contain adjacent intervals with opposite effects, making it difficult to derive a single coherent rule describing the pattern. Patterns defined by simple magnitude effects (e.g., consistently high or low values) can be partially inferred from the rules, since threshold conditions with strong positive weights often align with the injected feature behaviour. In contrast, correlation-based patterns cannot be directly extracted because LIME expresses explanations as independent univariate conditions and does not encode relationships between features. Similarly, 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. As a result, pattern information becomes fragmented across multiple locally valid explanations, requiring substantial manual interpretation to reconstruct the intended structure. Overall, although LIME reliably highlights influential pattern features and provides insight into value ranges supporting or opposing predictions, the combination of fragmented rules and absence of explicit feature relationships makes it difficult to recover complete and interpretable global patterns from the explanations. TABLE I: Example results from LIME method for synthetic dataset with 50 features. Pattern Rule 1 Weight Rule 2 Weight Rule 3 Weight Rule 4 Weight Positive O1 H0 O1 <= -0.25 0.2289 H0 <= -0.26 0.162 -0.25 <O1 <= 0.11 -0.1015 0.11 <O1 <= 0.45 -0.0952 Negative B0 V1 V1 <= -0.19 0.1994 B0 <= -0.17 0.1978 -0.17 <B0 <= 0.13 -0.0868 0.13 <B0 <= 0.43 -0.0837 High R0 P1 R0 >0.51 0.0959 P0 >0.51 0.0734 -0.54 <R0 <= -0.03 -0.055 P0 <= -0.53 0.0339 Low O0 I0 I0 <= -0.43 0.2081 O0 <= -0.43 0.1344 -0.43 <I0 <= 0.09 -0.0714 0.09 <I0 <= 0.58 -0.0641 Switch V1 G0 G0 <= -0.46 0.1729 -0.46 <G0 <= 0.04 -0.0815 V1 >0.46 0.0622 -0.19 <V1 <= 0.16 -0.0427 Positive R1 V1 I1 R1 <= -0.24 0.2151 I1 <= -0.22 0.169 0.09 <R1 <= 0.43 -0.1038 -0.22 <I1 <= 0.14 -0.0826 Negative F0 L1 B1 F0 <= -0.27 0.2125 B1 <= -0.25 0.212 L1 <= -0.28 0.2035 -0.25 <B1 <= 0.09 -0.0892 High C0 B0 F0 F0 >0.44 0.0649 C0 >0.53 0.0588 -0.27 <F0 <= 0.06 -0.05 -0.05 <C0 <= 0.53 -0.0449 Low N0 O0 G1 N0 <= -0.45 0.2002 G1 <= -0.43 0.197 O0 <= -0.43 0.0723 0.08 <G1 <= 0.59 -0.0681 Positive M1 B0 J1 M0 J1 <= -0.24 0.1488 M0 <= -0.26 0.1306 M1 <= -0.25 0.1281 -0.24 <J1 <= 0.12 -0.0685 High-Low T1 P1 S1 I1 S1 <= -0.44 0.1492 I1 <= -0.22 0.1203 P1 <= -0.51 0.0571 -0.44 <S1 <= 0.11 -0.0528 Switch A1 O0 C1 L1 C1 <= -0.49 0.1488 A1 <= -0.46 0.1463 -0.49 <C1 <= 0.05 -0.0562 O0 <= -0.43 0.0456 VI Conclusions and Future Work Using synthetic datasets with injected ground-truth patterns, this work evaluates RF surrogate models, PCA and LIME in terms of their ability to recover relevant features and provide interpretable representations of underlying structures. The results demonstrate that, although 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. The RF surrogate model showed strong and consistent ranking performance across pattern types and data dimensionalities, successfully prioritising relevant features. However, its outputs were limited to feature importance rankings, providing little insight into how features relate to one another or collectively form patterns. 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. Consequently, PCA proved sensitive to the statistical structure of the data and unreliable for detecting more complex or subtle patterns. LIME, while offering rule-based explanations and successfully identifying influential features, produced fragmented valid rules that did not coherently represent global feature relationships. Correlation-based and switching patterns, in particular, could not be directly recovered due to the method’s reliance on independent univariate conditions. Taken together, these findings highlight a broader methodological limitation: current widely used techniques are primarily designed for feature ranking, dimensionality reduction, or local explanation rather than true pattern detection. As a result, identifying structured multi-feature behaviours remains challenging, even in controlled settings where patterns are known to exist. This limitation 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 lack of dedicated pattern detection methodologies therefore represents a significant gap in current data analysis practice. 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. Evaluation of existing pattern detection methods against real-world datasets was beyond the scope of this work. Future work should expand this evaluation to real-world datasets. Whilst testing on synthetic datasets provides a controlled experimental environment, evaluation of existing approaches’ efficacy identifying known and documented patterns in real-world datasets would improve confidence in the evaluation. This would also provide a comparative baseline for future pattern detection methods to be evaluated against, further improving confidence in novel approaches. VII Acknowledgements Benjamin Connor is supported by a Department for the Economy Research Studentship. References [1] M. Alvarez-Garcia, R. Ibar-Alonso, and M. Arenas-Parra (2024) A comprehensive framework for explainable cluster analysis. Information Sciences 663, p. 120282. Cited by: §I. [2] S. Bandyapadhyay, F. V. Fomin, P. A. Golovach, W. Lochet, N. Purohit, and K. Simonov (2023) How to find a good explanation for clustering?. Artificial Intelligence 322, p. 103948. Cited by: §I, §I. [3] V. Belle and I. Papantonis (2021) Principles and practice of explainable machine learning. Frontiers in big Data 4, p. 688969. Cited by: §I-A. [4] H. Dunn, L. Quinn, S. J. Corbridge, K. Eldeirawi, M. Kapella, and E. G. Collins (2018) Cluster analysis in nursing research: an introduction, historical perspective, and future directions. Western journal of nursing research 40 (11), p. 1658–1676. Cited by: §I. [5] A. K. Jain (2010) Data clustering: 50 years beyond k-means. Pattern recognition letters 31 (8), p. 651–666. Cited by: §I, §I. [6] I. T. Jolliffe and J. Cadima (2016) Principal component analysis: a review and recent developments. Philosophical transactions of the royal society A: Mathematical, Physical and Engineering Sciences 374 (2065), p. 20150202. Cited by: §I-C. [7] T. J. Loftus, B. Shickel, J. A. Balch, P. J. Tighe, K. L. Abbott, B. Fazzone, E. M. Anderson, J. Rozowsky, T. Ozrazgat-Baslanti, Y. Ren, et al. (2022) Phenotype clustering in health care: a narrative review for clinicians. Frontiers in artificial intelligence 5, p. 842306. Cited by: §I, §I, §I. [8] R. Luss and A. d’Aspremont (2010) Clustering and feature selection using sparse principal component analysis. Optimization and Engineering 11 (1), p. 145–157. Cited by: §I. [9] G. Marín Díaz, R. Gómez Medina, and J. A. Aijón Jiménez (2024) Integrating fuzzy c-means clustering and explainable ai for robust galaxy classification. Mathematics 12 (18), p. 2797. Cited by: §I. [10] M. Moshkovitz, S. Dasgupta, C. Rashtchian, and N. Frost (2020) Explainable k-means and k-medians clustering. In International conference on machine learning, p. 7055–7065. Cited by: §I. [11] M. T. Ribeiro, S. Singh, and C. Guestrin (2016) ” Why should i trust you?” explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, p. 1135–1144. Cited by: §I-B. [12] E. Werner, J. N. Clark, A. Hepburn, R. S. Bhamber, M. Ambler, C. P. Bourdeaux, C. J. McWilliams, and R. Santos-Rodriguez (2023) Explainable hierarchical clustering for patient subtyping and risk prediction. Experimental Biology and Medicine 248 (24), p. 2547–2559. Cited by: §I. [13] E. Yamga, L. Mullie, M. Durand, A. Cadrin-Chenevert, A. Tang, E. Montagnon, C. Chartrand-Lefebvre, and M. Chassé (2023) Interpretable clinical phenotypes among patients hospitalized with covid-19 using cluster analysis. Frontiers in Digital Health 5, p. 1142822. Cited by: §I. [14] M. R. Zafar and N. M. Khan (2019) DLIME: a deterministic local interpretable model-agnostic explanations approach for computer-aided diagnosis systems. arXiv preprint arXiv:1906.10263. Cited by: §I.