Objective Similarity Measures

• Objective similarity measures are mathematical functions that quantify the resemblance between objects using formal properties like symmetry, reflexivity, and boundedness.
• They include methods such as Euclidean, cosine, and Jaccard measures, which are applied across vectors, sets, graphs, and semantic spaces for diverse data types.
• These measures can be composed, learned from data, or benchmarked against standardized tasks, offering flexibility in applications from machine learning to network analysis.

Objective similarity measures are mathematical functions designed to quantify the degree to which two objects, signals, patterns, sets, or representations are alike according to well-defined, non-arbitrary criteria. These measures are defined by formal properties—such as symmetry, reflexivity, boundedness, and (in some cases) metricity—and typically eschew subjective parameter choices or hand-tuned heuristics. Objective similarity can be established axiomatically, learned from empirical data, or defined through domain-specific structural or statistical criteria, providing a standardized backbone for tasks in machine learning, information retrieval, network science, natural language processing, and beyond.

1. Mathematical Foundations of Objective Similarity

The formal approach to objective similarity and dissimilarity starts with a function on a pair of objects $(x,y)$ from a space $X$. A similarity measure is defined as $s: X \times X \to I_s \subset \mathbb{R}$, which is total, exhaustive, and upper-bounded; its “dual” dissimilarity (or distance) measure $d: X \times X \to I_d \subset \mathbb{R}$ is lower-bounded. Key properties include:

• Reflexivity: $s(x,x) = S_{\max}$; $d(x,x) = d_{\min}$
• Symmetry: $s(x,y) = s(y,x)$; $d(x,y) = d(y,x)$
• Strong reflexivity, boundedness, closedness, complementarity, and transitivity via an operator $T$ (e.g., triangle inequality for distances).
• Duality: Any strictly decreasing bijection $f$ maps a similarity to a dissimilarity via $X$0, preserving key properties under the transformation (Belanche, 2012).

For practical and theoretical flexibility, similarity/distance functions can be composed with increasing or decreasing bijections, yielding equivalence classes of measures with the same ordering of object pairs.

2. Classical Vector and Set-Based Similarity Measures

Objective similarity for vectors and sets is established through canonical functions:

• Distance Measures: Euclidean (L2), Manhattan (L1), Minkowski, Chebyshev, and Mahalanobis distances are all metrics (except in specific degenerate parameterizations), with standard properties and complexity $X$1 per pair. Each metric emphasizes different aspects of difference—overall magnitude, coordinatewise shifts, outlier sensitivity, and scale/correlation invariance (levy et al., 2024).
• Inner-Product and Cosine Similarity: $X$2 is widely used, especially for high-dimensional or sparse data where normalization is crucial. Cosine distance is not a metric but underlies many kernel methods and information retrieval systems.
• Set/Histogram Overlap: Measures like Jaccard, Dice, and histogram intersection ($X$3) are grounded in set theory; Jaccard ($X$4) is a metric for sets and is a “smoothed” version of the Kronecker delta (Costa, 2021). These measures extend naturally to nonnegative vectors.
• Correlation-Based Measures: Pearson and Spearman correlation coefficients quantify linear or monotonic association, respectively, and are commonly converted to “distance” by $X$5, yet are not metrics (levy et al., 2024).
• Joint Measures (e.g., JDM): The Joint Distance Measure (Awotunde, 9 Apr 2025) combines spatial (Minkowski) and angular (cosine) information:

$X$6

with its own trade-off between magnitude and directionality; however, it is not a true metric due to failure of the triangle inequality.

3. Objective Similarity in Structured and Symbolic Domains

Objective measures in graphs, networks, and semantic spaces employ domain-specific formalizations:

• Networks/Graphs:
  • Vertex Similarity: Spectral simplex embeddings represent each vertex as a convex combination of archetypes; $X$7 distance in this mixture space is used for quantifying within-network similarity (Tsourakakis, 2011).
  • Graph Matching: The generalized condition number $X$8 between Laplacians of two graphs quantifies structural similarity and underpins deterministic and randomized alignment algorithms. $X$9 signals exact isomorphism (Tsourakakis, 2011).
  • Graph Embeddings: VERSE learns an embedding so that, for arbitrary node similarity distribution $s: X \times X \to I_s \subset \mathbb{R}$0, the softmax-induced embedding similarity $s: X \times X \to I_s \subset \mathbb{R}$1 minimizes the KL divergence $s: X \times X \to I_s \subset \mathbb{R}$2 (Tsitsulin et al., 2018).
• Semantic and Knowledge Spaces:
  • Ontology/Taxonomy Measures: Path-based (e.g., Wu–Palmer, Leacock–Chodorow), information–content (e.g., Lin, Resnik), and feature-based (e.g., Tversky set overlap) measures are built upon the structure and/or corpus statistics of taxa (Slimani, 2013, K et al., 2012). Hybrid measures interpolate these features for maximum empirical fit.
  • Edge-Counting with Penalization: To correct for “neighbor inflation” (artifacts of deep LCS or branchings), penalization factors based on path length and maximum taxonomy depth are used:

  $s: X \times X \to I_s \subset \mathbb{R}$3

  yielding empirically stronger correlation with human ratings (K et al., 2012). - Distributional Semantics: Cosine similarity on PPMI vectors remains standard but is challenged by intersection-based scores like APSyn, which sums reciprocals of mean ranks over overlapping salient contexts, delivering gains in “genuine similarity” detection over association (Santus et al., 2016).

4. Learning Objective Similarity from Data

Objective similarity can be learned directly from labeled data via supervised or self-supervised optimization:

• Pairwise or Triplet Learning: Given pairs or triplets with ground truth similarity/preference, measures such as deep metric learning, Siamese architectures, and triplet loss objectives are applied. For example, in large-scale crowdsourced material similarity, a deep ResNet with bespoke loss aligns embedding-space distances to forced-choice human judgments with both margin ranking and probabilistic likelihood components (Lagunas et al., 2019).
• Direct Supervised Regression: For texts, “STSScore” fine-tunes a regressor to predict mean human similarity directly from sentence pairs, vastly outperforming token overlap and embedding-cosine measures on STS-Benchmarks (Herbold, 2023).
• Case-Based Reasoning: Objective similarity measures can be constructed by learning from problem/solution pairs, fitting either a post hoc comparison function over learned embeddings, or directly leveraging classifier outputs as soft similarity functions (Mathisen et al., 2020).
• Data-Driven Local Similarity: For mixed-type tabular data, polynomial decay for numerical attributes and full-range categorical similarity ensure automated normalization and consistent use of the similarity interval 0,1.

5. Systematic Testing and Benchmarking of Similarity Measures

Comprehensive benchmarks such as ReSi evaluate objective representational similarity measures for their ability to consistently detect functional, geometric, or designed similarities under controlled settings (Klabunde et al., 2024):

• Grounding by Prediction: Correlation to test-set accuracy or output distribution differences among neural models.
• Grounding by Design: Group separation based on model perturbations—randomization, shortcut affinity, augmentations, and monotonicity of inner-layer similarity.
• Measure Taxonomy: Kernel alignment (CKA), RSM-based, alignment-based (e.g., Procrustes), neighbors-overlap (e.g., kNN Jaccard), CCA-based, statistical descriptors, and topological (Menger curvature) measures offer complementary performance.
• Empirical Guidance: No single measure dominates all settings; matching the similarity measure to the induction principle (e.g., functional similarity vs. group discrimination vs. local geometry) yields best results.

Class	Example Formula	Typical Data Type
L2, L1, Lp	$s: X \times X \to I_s \subset \mathbb{R}$4	Numeric vectors
Cosine similarity	$s: X \times X \to I_s \subset \mathbb{R}$5	Sparse/high-dim vectors
Jaccard	$s: X \times X \to I_s \subset \mathbb{R}$6	Sets, binary vectors
Information-content	Lin, Resnik, etc.	Ontology, taxonomy
APSyn	Eq. over salient context overlap	Distributional semantics
CKA	See (Klabunde et al., 2024)	Neural representations

6. Objective Similarity in Specialized Domains and Extensions

• Material Appearance: Objective learned metrics combine 2AFC crowdsourcing with deep feature learning, enabling direct application in retrieval and visualization tasks for appearance-based design (Lagunas et al., 2019).
• Soft Sets: Set-operation-based definitions correct earlier matrix-based similarity formulas for uncertainty spaces, achieving proper metricity and parameter-sensitivity (Kharal, 2010).
• Taxonomically Anchored Text Similarity: Cross-taxon semantic similarity is refined via corpus statistics or domain-specific constraints (e.g., Bloom's taxonomy for educational objectives) (Pawar et al., 2018).
• Complex Structured Inputs: Methodologies generalize from scalars to multisets, vectors, and functionals, recovering Jaccard and inner-product (cosine) metrics as special cases and supporting normalized, sign-sensitive, and convolutional extensions (Costa, 2021).

7. Practical Selection and Application Guidelines

• Data and domain type dictate the suitable objective similarity class: metric (distance-based) for geometric retrieval, inner-product-based for document or feature vectors, IC- or feature-based for ontologies, and supervised or learned measures wherever human labels or functional criteria are available (levy et al., 2024).
• Interpretability concerns favor sparsity and transparency—e.g., tf-idf cosine for explainable IR (Sitikhu et al., 2019).
• Metric properties (triangle inequality) are necessary for indexing and some algorithms, but not all similarity functions satisfy them (e.g., JDM).
• Benchmarking across groundings is essential for new application domains; always triangulate results using multiple measure classes (Klabunde et al., 2024).
• Automated or data-driven approaches (learned metrics, data-derived scaling) are preferred when labeled data is available and interpretability can be managed.

In sum, objective similarity measures constitute a rigorously grounded, empirically validated apparatus for quantifying resemblance across a spectrum of data types and scientific applications. They are defined by adherence to explicit algebraic and probabilistic properties, and can be evaluated, extended, or learned by matching the mathematical formalism to the intended scientific hypothesis or downstream task.