Element-Centric Similarity Overview

• Element-centric similarity is defined as a framework where global similarity is built from the local contributions and relationship profiles of individual elements.
• It employs diverse mechanisms such as coordinate-wise operations, spectral decomposition, and optimal transport to capture both granular differences and overall structure.
• This approach overcomes limitations of rigid cluster-centric measures, enhancing interpretability in applications like deep metric learning, clustering analysis, and materials representations.

Element-centric similarity denotes, across several research traditions, a mode of comparison in which similarity is determined from the contributions of individual elements or from each element’s relationship profile to the rest of a system, rather than only from global aggregates or cluster-level summaries. The elemental unit may be a feature coordinate in a learned representation, a member of a multiset, a data point in a similarity matrix, a DOM node, a layout object, a chemical element, or a vertex in a clustering-induced graph. Taken together, the cited works suggest that the unifying principle is local contribution followed by structured aggregation: similarity is built from element-wise absolute differences, products, overlaps, transport costs, ordinariness scores, eigenvalue spectra, or diffusion-based affinity profiles (Zhu et al., 2017, Costa, 2021, Nguyen et al., 5 Nov 2025, Otani et al., 2024, Nass et al., 2022, Onwuli et al., 2023, Gates et al., 2017).

1. Conceptual foundations

The literature suggests that element-centric similarity is best understood as a design principle rather than a single metric. In one class of formulations, each coordinate or element contributes directly to the final score; in another, each element is first assigned a neighborhood-dependent profile, and comparison is performed between those profiles. This produces a family of methods that remain local in construction even when the final similarity is global.

Domain	Elemental unit	Element-centric mechanism
Person re-identification	Feature dimension $i$	Learned weights on $|X_{1,i}-X_{2,i}|$ and $X_{1,i}X_{2,i}$
Sets, vectors, functions	Scalar entry or point $x$	Aggregation of min/max/product-based scalar similarities
Similarity-sensitive entropy	Unique element $i$ or observation	Ordinariness $(Z\mathbf p)_i$ or eigenvalue spectrum of $Z_n/n$
Layouts and web GUIs	Layout element or DOM node	Element-to-element transport cost or weighted multi-attribute score
Clustering comparison	Data element $v_i$	Personalized PageRank affinity profile induced by cluster structure
Materials representations	Chemical element	Distances, cosine similarity, or Pearson correlation in embedding space

A recurrent contrast in these papers is with more rigid or more global alternatives. Cluster-centric comparison measures operate on clusters or pairs of elements; strict layout metrics rely on one-to-one same-category matching; Shannon-type entropy ignores similarity structure; radius-ratio rules reduce chemistry to a low-dimensional heuristic; and single locators or exact XPaths treat localization as a brittle yes-or-no event. Element-centric methods replace these with graded local evidence and explicit aggregation (Gates et al., 2017).

2. Coordinate-wise and algebraic formulations

A fundamental template appears in scalar, vector, and feature-learning settings. In "On Similarity," the prototypical reference is the Kronecker delta, which equals 1 when two real values are identical and 0 otherwise. The paper then derives yielding indices that relax strict identity into graded similarity. For unsigned scalars, one core index is

$s_1(x,y)=\frac{2\min\{|x|,|y|\}}{|x|+|y|},$

with two further bounded variants,

$$s_2(x,y)=\frac{\min\{|x|,|y|\}}{\max\{|x|,|y|\}}, \qquad s_3(x,y)=\frac{|x|\,|y|}{(\max\{|x|,|y|\})^2},$$

and an unbounded product form $|X_{1,i}-X_{2,i}|$0. Signed generalizations incorporate $|X_{1,i}-X_{2,i}|$1, so opposition can produce negative similarity. The same scalar building blocks are then aggregated over multisets, vectors, and functions by summation or integration, yielding explicitly element-wise formulations. Within this framework, the Jaccard index is interpreted as a yielding implementation of the Kronecker delta, and the multiset index $|X_{1,i}-X_{2,i}|$2 generalizes Jaccard to real and negative multiplicities (Costa, 2021).

In deep metric learning, "Deep Hybrid Similarity Learning for Person Re-identification" gives a clean coordinate-wise instantiation. Two Siamese CNN branches produce feature vectors $|X_{1,i}-X_{2,i}|$3, after which the pair representation is built exclusively from element-wise operators: $|X_{1,i}-X_{2,i}|$4 The hybrid similarity is

$|X_{1,i}-X_{2,i}|$5

equivalently $|X_{1,i}-X_{2,i}|$6 with $|X_{1,i}-X_{2,i}|$7. Each dimension contributes

$|X_{1,i}-X_{2,i}|$8

so the final score is an additive sum of per-dimension contributions. The similarity layer therefore has only $|X_{1,i}-X_{2,i}|$9 parameters, rather than $X_{1,i}X_{2,i}$0 for a full Mahalanobis matrix. The paper explicitly interprets absolute difference as a distance-like cue and element-wise multiplication as a correlation-like cue; empirically, using both cues is more effective than either alone (Zhu et al., 2017).

These formulations establish a central mathematical theme of element-centric similarity: the score is decomposed into local comparisons that remain interpretable at the level of individual coordinates or scalar entries, while the overall model can still be optimized jointly with a larger system.

3. Similarity neighborhoods and entropy

A second major strand treats each element’s neighborhood in similarity space as the primitive object. "Which Similarity-Sensitive Entropy?" begins from the observation that Shannon, Rényi, Tsallis, and Hill measures depend only on frequencies $X_{1,i}X_{2,i}$1 and therefore ignore pairwise similarities among elements. To incorporate similarity, the paper uses a matrix $X_{1,i}X_{2,i}$2 with $X_{1,i}X_{2,i}$3. In the Leinster–Cobbold–Reeve framework, the ordinariness of element $X_{1,i}X_{2,i}$4 is

$X_{1,i}X_{2,i}$5

and the similarity-sensitive diversity is

$X_{1,i}X_{2,i}$6

This is element-centric by design: each element contributes according to its similarity-weighted neighborhood. Elements similar to many high-frequency elements are ordinary; elements dissimilar to almost everything are unusual.

The Vendi score uses the same similarity information in a different way. It is defined on the eigenvalues $X_{1,i}X_{2,i}$7 of the normalized PSD kernel $X_{1,i}X_{2,i}$8, with

$X_{1,i}X_{2,i}$9

and Rényi-type relatives

$x$0

The paper interprets these eigenvectors as “ur-elements” or “eigen-elements,” so VS is spectral rather than row-wise. The resulting contrast is conceptually sharp: LCR aggregates per-element ordinariness, whereas VS measures the effective number of occupied spectral directions.

A central practical parameter is the similarity scaling

$x$1

where $x$2 is described as a half distance. As $x$3, all similarities approach 1 and both LCR and VS approach 1; as $x$4, off-diagonal similarities vanish and both measures approach $x$5 when all elements are unique and equally weighted. At intermediate scales they can differ substantially. On a 10,000-image subset of MNIST, HDBSCAN found 6 clusters, LCR at $x$6 gave $x$7, and VS at $x$8 gave about $x$9. The paper concludes that VS is preferable only when interpreting elements as linear combinations of a more fundamental set of “ur-elements” or when the system has a quantum-mechanical character; in the broader circumstance of capturing information encoded by similarity, LCR is favored. It also proves that VS provides an upper bound on LCR for $i$0, proves $i$1 for full-rank $i$2 and any $i$3, and conjectures a general inequality $i$4 for all orders (Nguyen et al., 5 Nov 2025).

This branch of the literature shows that element-centric similarity is not limited to pairwise matching. It can also define uncertainty, redundancy, and effective number by making each element’s contribution depend on its neighborhood.

4. Structured objects, transport, and multi-attribute matching

For structured sets of elements, element-centric similarity often appears as flexible matching rather than exact correspondence. In the scalar-to-functional framework of "On Similarity," multisets, vectors, and functions are compared by aggregating element-wise overlaps based on min, max, and product operators. For functions $i$5, for example,

$i$6

and related functionals induce similarity-based convolutions and correlations. The common product functional $i$7 measures signed common area, while $i$8 plays a union-like role. The conceptual point is that global similarity is obtained by integrating point-wise elemental contributions, not by collapsing the functions to a few summary statistics (Costa, 2021).

"LTSim: Layout Transportation-based Similarity Measure for Evaluating Layout Generation" applies the same logic to 2D layouts represented as sets of labeled bounding boxes $i$9. Each layout is treated as a discrete probability distribution over its elements, and element-to-element dissimilarity is defined by

$(Z\mathbf p)_i$0

with $(Z\mathbf p)_i$1 and $(Z\mathbf p)_i$2 when labels match and 0 otherwise. Optimal transport then solves for a plan $(Z\mathbf p)_i$3 minimizing total cost under uniform marginals, producing $(Z\mathbf p)_i$4, and similarity is

$(Z\mathbf p)_i$5

Because transport is soft and many-to-many, one element can distribute mass across several counterparts, different cardinalities are handled naturally, and layouts sharing no categories can still be compared through geometry alone. At the collection level, the paper uses $(Z\mathbf p)_i$6 inside an unbiased empirical $(Z\mathbf p)_i$7, yielding LTSim-MMD as a distribution-level discrepancy (Otani et al., 2024).

"Similarity-based web element localization for robust test automation" transfers the same principle to DOM nodes. A target element and each candidate in the new DOM are represented by 14 locator parameters: Tag, Class, Name, Id, HRef, Alt, Absolute XPath, ID relative XPath, IsButton, Location, Area, Shape, Visible Text, and Neighbor Texts. Similo computes

$(Z\mathbf p)_i$8

using parameter-specific similarities such as case-insensitive equality, normalized Levenshtein distance, Euclidean-location decay with a 100-pixel cutoff, and word overlap for Neighbor Texts, then selects the arg max over all visible candidates. On 40 popular websites and 598 target elements, Similo failed in 72 cases, compared to 146 failed cases for the baseline multi-locator approach (Nass et al., 2022).

These methods reject strict one-to-one exact matching as the sole basis for similarity. Their common claim is that an object remains identifiable through a constellation of partial element-level cues even when exact locators, exact category matches, or identical counts are unavailable.

5. Element-induced relation profiles and clustering comparison

In clustering comparison, element-centric similarity is formulated explicitly as a contrast with cluster-centric measures. "Element-centric clustering comparison unifies overlaps and hierarchy" begins with a cluster affiliation graph $(Z\mathbf p)_i$9, a bipartite graph between elements $Z_n/n$0 and clusters $Z_n/n$1. The affiliation matrix $Z_n/n$2 has entries $Z_n/n$3, possibly weighted by a hierarchy function

$Z_n/n$4

when clusters occur at different dendrogram levels. Projecting the affiliation graph onto elements yields a cluster-induced element graph with weights

$Z_n/n$5

This weight encodes how strongly element $Z_n/n$6 is related to element $Z_n/n$7 through shared cluster membership, normalized by cluster size and membership multiplicity.

For each element $Z_n/n$8, the method then computes a personalized PageRank affinity vector

$Z_n/n$9

with $v_i$0. Element-wise similarity between clusterings $v_i$1 and $v_i$2 is

$v_i$3

and global similarity is the average

$v_i$4

The method applies uniformly to disjoint, overlapping, and hierarchical clusterings because all structure is encoded at the element–cluster edge level before diffusion.

The paper argues that standard measures suffer from critical biases. In synthetic experiments, NMI increases with the number of clusters in a random-equal-cluster setting, Jaccard and F increase with skewness in cluster-size distributions, and ARI and ONMI can collapse to zero in scenarios that retain nontrivial structural similarity. The element-centric framework is presented as avoiding these pathologies while also providing diagnostics unavailable to scalar cluster-centric scores. Because $v_i$5 is computed for each element, disagreement can be localized to individual data points, social-network vertices, or brain regions. The paper illustrates this with K-means on synthetic and handwritten-digit data, with Facebook friendship networks where the alignment of communities with class year and major differs across student cohorts, and with overlapping and hierarchical communities in resting-state fMRI, where element-centric similarity on the full hierarchy gives about 84% average classification accuracy for schizophrenia versus control and highlights the Fusiform gyrus as especially variable in patients (Gates et al., 2017).

This formulation extends the meaning of element-centric similarity beyond direct comparison of object parts. Here, an element is characterized by the relational world induced around it by a clustering, and similarity compares those induced worlds.

6. Representation spaces, applications, and methodological issues

Element-centric similarity also appears in scientific representation learning. "Element similarity in high-dimensional materials representations" represents each chemical element $v_i$6 by a vector $v_i$7 and defines similarity through distances or angular correlation: $v_i$8 The paper compares Magpie (22 dimensions), Mat2vec (200), MatScholar (200), MEGNet16 (16), Oliynyk (44), SkipAtom (200), and Random_200. PCA and t-SNE reveal clustering of lanthanoids, alkaline-earth groupings such as Sr–Ba–Ca, and distributed-representation similarities such as Bi–Sb. For binary AB crystal-structure prediction, cosine similarity is used inside a substitution model

$v_i$9

and on a 100-compound dataset the reported accuracies are 54% for radius ratios, 54% for Hautier weights, 58% for Random_200, 68% for SkipAtom, 73% for MEGNet16, 75% for Oliynyk, 78% for Magpie, 80% for Mat2vec, and 81% for MatScholar (Onwuli et al., 2023).

Across the cited literature, several methodological issues recur. Some formulations require strong structural assumptions: VS assumes a PSD similarity matrix, whereas LCR does not (Nguyen et al., 5 Nov 2025). Some methods use fixed heuristic weights or scales: Similo uses manually chosen parameter weights, LTSim fixes equal weighting between geometry and label, DHSL learns per-dimension weights but fixes the element-wise operator family, and the entropy literature shows that the half distance $s_1(x,y)=\frac{2\min\{|x|,|y|\}}{|x|+|y|},$0 can alter magnitudes by orders of magnitude (Nass et al., 2022, Otani et al., 2024, Zhu et al., 2017, Nguyen et al., 5 Nov 2025). Computational burden is another theme: VS requires eigen-decomposition of an $s_1(x,y)=\frac{2\min\{|x|,|y|\}}{|x|+|y|},$1 matrix over all observations, LTSim-MMD requires many pairwise OT computations, and clustering comparison may require repeated personalized PageRank evaluations, although analytic simplifications and approximate algorithms are discussed (Nguyen et al., 5 Nov 2025, Otani et al., 2024, Gates et al., 2017).

A common misconception is that element-centric similarity necessarily means exact element matching. The cited work instead shows several distinct possibilities: direct coordinate-wise addition, overlap/union aggregation, neighborhood ordinariness, spectral decomposition, optimal transport, diffusion on induced graphs, and similarity in learned embedding spaces. Another misconception is that element-centric methods are always more local and therefore less expressive. The diffusion-based clustering framework and the spectral entropy framework show the opposite: local element-level definitions can induce highly nonlocal global behavior (Gates et al., 2017, Nguyen et al., 5 Nov 2025).

In this sense, element-centric similarity is a general research program for building global comparison from local structure. Its defining move is to make the individual element, coordinate, or node the primary carrier of information, and then to design an aggregation mechanism that preserves the geometry, redundancy, or relational role of those elemental contributions.