Cluster-Feature Mapping Overview

• Cluster-feature mapping is a framework that directly relates data clusters to features using assignments, probability functions, or explicit mappings.
• It integrates statistical, deep learning, and Bayesian approaches to enhance interpretability, feature selection, and clustering performance.
• Practical applications span model-based feature selection, deep activation mapping, and scalable clustering in high-dimensional or distributed settings.

A cluster-feature mapping is any structure, algorithm, or mathematical formalism that specifies or computes a direct relationship between data clusters and features. This relationship can be represented as an assignment, probability function, or explicit mapping that indicates which features drive (or best distinguish) each cluster—or, conversely, how clusters can be understood as special combinations or allocations of features. Such mappings arise in a broad range of modern statistical learning, deep learning, and Bayesian nonparametrics, informing feature selection, interpretability, distributed clustering, deep representation learning, and combinatorial stochastic process models.

1. Mathematical Foundations: Clusters, Features, and Mappings

Cluster-feature mapping can be formalized through several statistical frameworks:

• Partitions vs. Feature Allocations: In traditional clustering, a partition of $[N]=\{1,\ldots,N\}$ is a set of mutually disjoint, exhaustive blocks. Each data point belongs to exactly one cluster. In feature allocation, each data point may belong to an arbitrary finite number of groups ("features"), and the assignment is represented as a binary matrix $F\in\{0,1\}^{N\times K}$, with $F_{nk}=1$ if data point $n$ has feature $k$.
• Exchangeable Probability Functions: The clustering context is modeled by exchangeable partition probability functions (EPPFs); the overlapping-feature case is handled using exchangeable feature probability functions (EFPFs). These mechanisms, and their paintbox representations, delineate the mathematical relations between clusters and features (see (Broderick et al., 2013, Broderick et al., 2012)).
• Mappings: There is a direct structural mapping: hard cluster assignment (partitions) are a special case of feature allocations where each row of $F$ sums to one. More general feature allocations allow cluster-feature relationships where clusters overlap in feature space and membership is not exclusive (Broderick et al., 2013, Broderick et al., 2012).

2. Model-Based Approaches: Cluster-Specific Feature Selection

Modern clustering algorithms frequently integrate feature selection or importance mapping to enhance interpretability and statistical power, especially in high-dimensional or heterogeneous settings.

• Nonparametric MAP Clustering with Per-Cluster Feature Selection: CRAFT formulates a joint clustering and feature-indicator model where each cluster $k$ is associated with a binary selection vector $v_{k,d}$ for feature $d$ (Garg et al., 2015). The assembled mapping at convergence directly specifies for each cluster $k$ the subspace $F\in\{0,1\}^{N\times K}$0—yielding explicit cluster-feature mappings that highlight which features are relevant for each cluster.
• Feature Screening: Convex fusion-based methods such as COSCI score each feature for its ability to induce large, balanced merges in a clustering tree, providing a ranking of features by cluster-leverage (Banerjee et al., 2017). Features above a threshold are retained, and subsequent clustering operates in this reduced feature space.

3. Deep Learning and Interpretability: Cluster Activation Maps

Deep neural architectures increasingly use explicit cluster-feature mapping for interpretability and robust feature attribution.

• CF-CAM: The Cluster Filter Class Activation Mapping pipeline applies density-based spatial clustering (DBSCAN) to the vectorized channels of a layer's activations, forming semantically cohesive channel clusters. Hierarchical importance weighting and gradient filtering yield CAMs with improved faithfulness and robustness, where the cluster-to-channel mapping explicitly reveals which activation patterns drive each decision (He et al., 31 Mar 2025).
• CLAM: In unsupervised deep clustering, CLAM constructs heatmaps identifying which spatial or feature regions determine cluster assignments, using channel-wise confidence scores and aggregation in a Score-CAM–adapted framework (Ryan et al., 2020).
• Neural Clustering Frameworks: Newer architectures such as FEC treat each feed-forward layer as an iteration of explicit clustering, with assignment matrices $F\in\{0,1\}^{N\times K}$1 giving fully transparent, human-interpretable mappings from pixels or patches to clusters at every depth (Chen et al., 2024).

4. Feature Allocation, Bayesian Nonparametrics, and Theoretical Unification

Bayesian nonparametric models provide a principled account of cluster-feature mappings, unifying clustering (partitions) and overlapping feature allocations under combinatorial stochastic process theory.

Model	Object	Exchangeable Law	Feature–Cluster Mapping
Dirichlet Process/CRP	Partition	EPPF	Each data point assigned to one cluster (partition)
Beta Process/IBP	Feature allocation	EFPF	Each data point assigned to any subset of features (overlap allowed)

• Paintbox Construction: Every exchangeable partition (clustering) can be represented by Kingman's paintbox; every exchangeable feature allocation corresponds to a feature paintbox—a collection of subsets of $F\in\{0,1\}^{N\times K}$2 with overlapping structure (Broderick et al., 2013). Feature-frequency models (FFMs) correspond to product-Bernoulli allocations and fully characterize EFPF models up to Poissonian noise (Broderick et al., 2013, Broderick et al., 2012).
• Mapping Reductions: By constraining feature allocations to one feature per data point, EFPF reduces to EPPF; conversely, relaxing disjointness in EPPF yields models supporting overlapping substructure (Broderick et al., 2013).

5. Practical Algorithms and Pipelines Leveraging Cluster–Feature Mapping

A variety of practical pipelines implement explicit or implicit cluster-feature mapping for efficient computation, interpretability, or domain-informed analysis.

• Feature Selection with Cluster Preservation: FSMLP_struct designs loss functions that penalize deviations in pairwise distances (Sammon stress) between the original data and the space of selected features, enforcing the preservation of cluster geometry alongside class discrimination (Das et al., 2023).
• Ensemble Feature Selection via Clustering: In domains with highly correlated or redundant features, pre-clustering of features followed by representative selection per cluster (e.g., by prognostic power in Cox models), then feature selection ensembles, improves stability and interpretability in domains such as Alzheimer's disease biomarker discovery (Spooner et al., 2022).
• Deep Clustering with Cluster-Feature Learnability: Learning paradigms such as FPGC equip each node (or data point) with a personalized “cluster-relevant feature set” via squeeze-and-excitation blocks, yielding, for each cluster, the union of the most predictive feature dimensions across members (Xie et al., 2024).
• Distributed/Privacy-Preserving Clustering with Overlapping Features: In federated or distributed settings with partial overlaps of feature spaces across sites, cluster-feature mapping is implemented through merging locally computed cluster statistics and reconciling global centroids via coordinate-wise aggregation (Maritan et al., 10 Oct 2025).

6. Interpretability, Transfer, and Generalization

The cluster-feature mapping paradigm has direct implications for model interpretability, transfer learning, and domain adaptation.

• Interpretability: CF-CAM and CLAM-type models provide saliency maps or region-attribute attributions for visual, clinical, or scientific inspection (He et al., 31 Mar 2025, Ryan et al., 2020).
• Generalization via Proxy Compression: Cluster-based proxy-labeling (e.g., ClusterFit) reduces overfitting in self-supervised representations and enhances transfer performance across tasks and modalities by smoothing away task-specific artifacts while retaining discriminative cluster structure (Yan et al., 2019).
• Structure-Preserving Feature Selection: Incorporating cluster preservation in feature selection (e.g., via geometric stress regularizers in FSMLP_struct) yields feature sets that simultaneously support high clustering performance and classification accuracy (Das et al., 2023).

7. Broader Impact and Unification: Theoretical and Applied Perspectives

Cluster-feature mapping serves as a unifying construct connecting several strands of research:

• Theoretical Unification: The paintbox and combinatorial stochastic process frameworks show that partition-based clustering and overlapping-feature allocation models can be embedded in a broader probabilistic machinery. Normalization and thinning connect Dirichlet and Beta processes, enabling transitions between clustering and feature allocation perspectives (Broderick et al., 2012, Broderick et al., 2013).
• Algorithmic Design: Cluster-feature mapping underlies many pipelines for scalable clustering, interpretable deep learning, privacy-preserving distributed learning, and domain-specific structure discovery—in genomic analysis, medical imaging, visual reasoning, and graph mining.

Cluster-feature mapping thus denotes a foundational concept and a suite of methods for representing, computing, and interpreting the relationships between clusters and features, with mathematical rigor and broad application from statistical theory to advanced neural architectures.