Evidential C-Means (ECM) Overview

• Evidential C-Means (ECM) is an evidential clustering method that assigns mass functions over power sets to capture uncertainty, ambiguity, and imprecision in class memberships.
• ECM generalizes fuzzy c-means by representing overlapping and ambiguous data points through credal partitions and barycentric prototype geometry.
• Advanced variants like MECM, TECM, and Soft-ECM extend ECM to relational, transfer, and non-Euclidean data settings by adapting prototype models and optimization criteria.

Evidential C-Means (ECM) is an evidential clustering method grounded in belief function theory, or Dempster–Shafer belief function theory, in which each object is assigned a mass function over the power set of classes rather than a membership vector over single classes alone. In this formulation, an object may support a single class, a subset of classes, or the empty set, and the resulting output is a credal partition rather than a hard or fuzzy partition. The method is commonly presented as a direct generalization of fuzzy c-means (FCM), but its distinguishing feature is not merely softened assignment: it is the explicit representation of imprecision, ambiguity, ignorance, and, in several formulations, outlierness or rejection within a single partitioning formalism (Jiao et al., 2021).

1. Belief-function basis and credal partitions

ECM is defined on a frame of discernment

$\Omega=\{\omega_1,\dots,\omega_c\},$

where $c$ is the number of classes. For each object, ECM estimates a basic belief assignment (bba), or mass function, on $2^\Omega$, the power set of $\Omega$. In the community-detection formulation, partial knowledge regarding the actual cluster of node $n_i$ is represented by a basic belief assignment $m$ from the power set of $\Omega$ to $[0,1]$, verifying

$\sum_{A \subseteq \Omega} m(A)=1.$

A subset $A \in 2^\Omega$ with $c$0 is called a focal element (Zhou et al., 2015).

This representation yields the credal partition. For a data set or graph with $c$1 objects, the credal partition is an $c$2-tuple

$c$3

where each $c$4 is a bba over $c$5. The central semantic distinction is that ECM permits mass on singletons, such as $c$6, and also on imprecise subsets, such as $c$7 or even $c$8. In the formulations summarized in later work, the empty set $c$9 is also used to represent atypicality, outliers, noise, or rejection (Jiao et al., 2021).

Belief and plausibility functions are the standard induced quantities: $2^\Omega$0

$2^\Omega$1

The pignistic transformation converts credal information into probabilities,

$2^\Omega$2

and is used when a crisp or fuzzy-like decision must be extracted from the credal partition (Zhou et al., 2015).

2. Optimization criterion and prototype geometry

For vector data $2^\Omega$3, ECM minimizes an objective over credal assignments and singleton cluster centers. One standard formulation is

$2^\Omega$4

subject to

$2^\Omega$5

Here $2^\Omega$6 is the cardinality of focal set $2^\Omega$7, $2^\Omega$8 penalizes focal sets with higher cardinality, $2^\Omega$9 is the fuzzifier or credibility exponent, and $\Omega$0 is the distance to the empty set, controlling outlier handling (Jiao et al., 2021).

The geometric mechanism of ECM is the barycentric construction of prototypes for imprecise focal sets. If $\Omega$1 denotes the singleton center of class $\Omega$2, then the barycenter associated with focal set $\Omega$3 is

$\Omega$4

Thus, precise classes have singleton prototypes, whereas imprecise subsets are represented by barycenters of the corresponding singleton centers (Jiao et al., 2021).

This geometry explains both the power and the restriction of the classical formulation. The first term in the objective pulls objects toward focal-set barycenters; the second allows mass to go to $\Omega$5 when an object is far from all clusters. At the same time, the method assumes that data are represented in a Euclidean space and that barycenters are meaningful there. Later reformulations treat this Euclidean dependence as a fundamental limitation for mixed, categorical, relational, and time-series data (Soubeiga et al., 17 Jul 2025).

3. Representation of overlap, uncertainty, ignorance, and outliers

ECM’s most distinctive property is that cluster membership is expressed over subsets of classes rather than single classes only. An assignment $\Omega$6 indicates support for a precise class; $\Omega$7 indicates ambiguity between two classes; $\Omega$8 indicates complete ignorance; and, in several formulations, $\Omega$9 indicates that the object is atypical or not explainable by the modeled classes (Jiao et al., 2021).

This subset-valued representation is not equivalent to ordinary fuzzy overlap. In the community-detection setting, overlap is explicit at the level of focal sets: a node is overlapping if ECM assigns significant mass to a subset containing multiple communities. The network paper states, for example, that in the Karate Club example “for ECM, node 1,9,10,12,31 belong to two clusters at the same time,” and in the Twitter example two nodes are partitioned into overlap communities $n_i$0 and $n_i$1 (Zhou et al., 2015).

A recurring misconception is to identify ECM with fuzzy c-means plus a post hoc threshold. The cited community-detection work contrasts ECM with fuzzy methods that require a threshold $n_i$2 to convert fuzzy membership into a final overlapping structure and notes that “there is no criterion for determining the appropriate $n_i$3.” By contrast, ECM represents overlap and uncertainty directly through focal sets and bbas, without requiring thresholding for that purpose (Zhou et al., 2015).

The same semantics also distinguishes uncertainty from ignorance more explicitly than conventional fuzzy partitions. In the median variant, this distinction is emphasized as one of the main advantages over fuzzy clustering, where equal memberships may conflate uncertainty and ignorance. A plausible implication is that ECM and its descendants are especially useful when boundary objects, bridge nodes, or structurally ambiguous observations should not be forced into over-specific decisions (Zhou et al., 2015).

4. ECM in spectral and graph-based community detection

In one influential graph-clustering use case, ECM is not applied directly to the graph. Instead, the method combines spectral mapping, ECM, and an evidential modularity criterion to obtain overlapping communities and select the number of classes (Zhou et al., 2015).

The graph is represented by an adjacency matrix $n_i$4 and a degree matrix $n_i$5 with $n_i$6. Spectral mapping uses the generalized eigensystem

$n_i$7

For $n_i$8, where $n_i$9 is an upper bound on the number of communities, the method finds the top $m$0 generalized eigenvectors $m$1, then forms the embedding $m$2. Each row of this matrix is treated as a sample in a $m$3-dimensional Euclidean space, and ECM is used to partition these samples into $m$4 classes. The output is a credal partition $m$5 for the graph (Zhou et al., 2015).

The number of communities is then selected by maximizing the evidential modularity

$m$6

which can also be written as

$m$7

Here $m$8 is the contour function associated to $m$9, and $\Omega$0 is the upper support for node $\Omega$1 belonging to community $\Omega$2. The paper characterizes $\Omega$3 as a direct extension of crisp modularity and states that when the credal partition degrades into a hard one, $\Omega$4 equals the ordinary crisp modularity $\Omega$5 (Zhou et al., 2015).

This pipeline makes ECM the final clustering engine operating on the spectral embedding, while evidential modularity is the model-selection criterion. The credal partition is presented as providing deeper insight into graph structure than crisp or fuzzy partitions because it can isolate nodes whose community membership is intrinsically imprecise (Zhou et al., 2015).

5. Relation to hard, fuzzy, possibilistic, and rough partitions, and practical limitations

ECM is described as generalizing hard, fuzzy, possibilistic, and rough partitions because it can encode precise membership, imprecise subset membership, complete ignorance, and empty-set assignment within one formalism (Jiao et al., 2021). Hard partitions correspond to the case where all mass is concentrated on one singleton. Fuzzy-like behavior arises when masses are restricted to singletons but remain fractional. Rough and possibilistic interpretations are likewise recovered through restrictions on the allowed focal sets and the semantics of uncertainty (Jiao et al., 2021).

This generality comes with statistical and computational cost. The transfer-learning study emphasizes that ECM must estimate numerous additional parameters compared with other partition-based algorithms, and therefore insufficient or contaminated data have a greater influence on its clustering performance. The paper contrasts the rough parameter burden schematically: c-means estimates $\Omega$6 assignments, FCM estimates roughly $\Omega$7 memberships, whereas ECM potentially estimates $\Omega$8 evidential masses. The same work states a complexity of

$\Omega$9

and notes that restricting focal sets to size at most $[0,1]$0 reduces this to

$[0,1]$1

Only a local optimum is guaranteed under the alternating optimization scheme (Jiao et al., 2021).

A second limitation concerns geometry. The classical ECM formulation assumes Euclidean data, Euclidean distances, and barycentric construction of imprecise-cluster prototypes. The Soft-ECM paper identifies this as the reason why existing evidential clustering algorithms cannot, in their original form, be applied to mixed data or non-tabular data such as time series, because such data are not generally represented in a Euclidean space and the classical algorithms rely on Euclidean barycenters (Soubeiga et al., 17 Jul 2025).

A third misconception concerns graphs. In the spectral community-detection framework, ECM is usable only after the graph has been embedded in Euclidean coordinates; it is not directly clustering the graph topology itself. By contrast, the median evidential c-means formulation removes the metric-space embedding requirement by constraining prototypes to belong to the original data set and using pairwise dissimilarities only (Zhou et al., 2015).

6. Median, transfer, and semi-metric generalizations

Several later methods preserve ECM’s evidential semantics while modifying its prototype model or optimization criterion.

Median Evidential C-Means (MECM) is a median or relational generalization in which singleton prototypes are constrained to be actual data objects,

$[0,1]$2

MECM keeps the ECM-style evidential objective and mass-update structure, but replaces Euclidean distances to barycenters by a relational or median dissimilarity $[0,1]$3. This removes the need for metric-space embedding and makes the method applicable to graph clustering and community detection. In the graph setting, MECM is coupled with an initialization strategy based on evidential semi-centrality and with evidential modularity for selecting the number of communities (Zhou et al., 2015).

Transfer Learning-based Evidential C-Means (TECM) addresses the sensitivity of ECM to insufficient or contaminated target data by integrating knowledge learned from a source domain into evidential clustering of the target domain. The transfer mechanism operates through learned source barycenters and an association matrix between source and target barycenters; the method is stated to be applicable when the source and target domains have the same or different numbers of clusters. The paper reports that TECM consistently outperforms ECM and gives average improvements over ECM of about $[0,1]$4 for insufficient target data and $[0,1]$5 for contaminated target data, while retaining evidential uncertainty modeling (Jiao et al., 2021).

Soft-ECM reformulates ECM for complex data in an abstract space $[0,1]$6 equipped with a semi-metric $[0,1]$7, rather than a Euclidean metric. Instead of defining imprecise-cluster prototypes as arithmetic barycenters, it treats them as optimization variables and introduces a relaxed objective with a consistency term weighted by $[0,1]$8. The paper states that if $[0,1]$9 and $\sum_{A \subseteq \Omega} m(A)=1.$0 is squared Euclidean distance, then Soft-ECM and ECM are equivalent; otherwise, the reformulation extends ECM to mixed data, categorical data, and time series, including settings that use Hamming distance or Soft-DTW (Soubeiga et al., 17 Jul 2025).

Taken together, these variants preserve the core ECM principle: clustering by bbas over $\sum_{A \subseteq \Omega} m(A)=1.$1 so that precise classes, imprecise meta-classes, and empty-set assignments remain part of the same partitioning language. What changes across MECM, TECM, and Soft-ECM is not the evidential semantics, but the geometry, optimization, and data assumptions under which that semantics can be made operational.