Neighbour-Homogeneity Consistency Score

• Neighbour-Homogeneity Consistency Score is a diagnostic metric that assesses whether local observations conform to an expected homogeneous behavior across varied data structures.
• It integrates methods such as Frobenius discrepancy in contingency tables, homophily ratios in graphs, and neighbor-difference penalties in regression to evaluate local model adequacy.
• The framework guides model calibration and weighting by comparing observed local compositions against simulated or theoretically derived homogeneous expectations.

“Neighbour‑Homogeneity Consistency Score” (Editor’s term) denotes a family of quantities that assess whether observations are locally consistent with a homogeneity assumption, where “local” may refer to adjacent cells in a contingency table, graph neighborhoods, nearest neighbors in covariate space, or adjacent indices in an ordered regression coefficient sequence. Across these settings, the common structure is the comparison of observed local composition with a model of homogeneous behavior: equal row distributions across columns in contingency tables, label or feature coherence in graph neighborhoods, outcome similarity among nearest neighbors, or equality of adjacent coefficients after data‑driven ordering. The term is not standardised across the cited literature, but the underlying idea appears repeatedly in forms such as Frobenius discrepancy for homogeneity testing, neighborhood homophily ratios, graph neighborhood congener ratios, 2‑hop neighbor class similarity, and neighbour‑difference penalties (Tygert, 2012, Peng et al., 17 Dec 2025, Gong et al., 2023, Cavallo et al., 2022, Sharpnack, 2019, Ke et al., 2013).

1. Conceptual definition and problem setting

A neighbour‑homogeneity consistency score is a local or semi‑local diagnostic of whether neighboring units behave as if they were generated by a common distribution, class mechanism, or parameter value. This suggests a unifying abstraction with three recurring components: a neighborhood system, an expected homogeneous pattern within that neighborhood, and a discrepancy or consistency functional comparing observed and expected behavior.

In contingency tables, the neighborhood is not spatial but structural: the homogeneity model fixes column totals and assumes a common row distribution across columns. In graph learning, the neighborhood is explicit through adjacency or $k$-hop reachability. In nearest neighbor matching, neighborhoods are induced by Voronoi cells and nearest‑neighbor relations in covariate space. In regression with homogeneity of coefficients, neighborhoods arise from adjacency in an ordering $\tau$ or from within‑segment and between‑segment relations in ordered segmentation (Tygert, 2012, Sharpnack, 2019, Ke et al., 2013).

The literature also distinguishes different semantic meanings of “homogeneity.” One meaning is equality of conditional distributions across groups, as in contingency tables. A second is label agreement or label purity in graph neighborhoods. A third is representativeness: nearby samples being sufficiently similar that substitution or averaging is asymptotically valid. A fourth is equality of parameters across neighboring indices or clustered covariates. A plausible implication is that “neighbour‑homogeneity consistency” is not a single statistic but a design pattern for constructing statistics tailored to the local structure of the data domain.

This pattern yields several scales of analysis. In graph learning, the quantities can be node‑wise, edge‑wise, or graph‑level. In regression, they can be pairwise between neighboring coefficients, segment‑level, or global through partition recovery. In contingency tables, the same homogeneity model can be interrogated globally with a single discrepancy statistic, but the construction can be extended conceptually to local neighborhoods by replacing global expected counts with local expectations based on neighboring cells (Peng et al., 17 Dec 2025, Ke et al., 2013, Tygert, 2012).

2. Homogeneity testing in contingency tables

The most direct statistical antecedent is the model for homogeneity of proportions in an $r \times s$ contingency table. With cell counts $n_{j,k}$, row totals $n_{j,}$, column totals $n_{,k}$, and grand total $n$, the model fixes the column totals and treats each column as a multinomial draw from a common row distribution $(p_1,\dots,p_r)$, with

$$(N_{1,k},\dots,N_{r,k}) \sim \text{Multinomial}\left(n_{,k}; p_1,\dots,p_r\right), \quad k = 1,\dots,s,$$

and

$$p_j = \frac{n_{j,}}{n}, \quad e_{j,k} = \frac{n_{j,} \cdot n_{,k}}{n}.$$

The expected counts are therefore identical to those of the standard independence model; the difference lies in the conditioning and interpretation of the sampling process (Tygert, 2012).

The central discrepancy in this setting is the Frobenius distance

$\tau$0

with random analogue

$\tau$1

and Monte Carlo $\tau$2-value

$\tau$3

This is an unscaled Euclidean discrepancy on the residual matrix. Tygert compares it with Pearson chi‑square, log‑likelihood ratio $\tau$4, and Freeman–Tukey/Hellinger distance, all computed from the same expected counts $\tau$5 but with different variance scalings or transforms (Tygert, 2012).

Several empirical examples in the paper show markedly smaller Monte Carlo $\tau$6-values for $\tau$7 than for classical statistics. In the Danish election polls example, $\tau$8, whereas $\tau$9, $r \times s$0, and $r \times s$1. In “Maniacal patients: reasons for termination,” $r \times s$2 versus $r \times s$3, $r \times s$4, and $r \times s$5. The paper also notes a counterexample, “Maniacal patients: reaction to prior lithium,” where the smallest $r \times s$6-value is Hellinger rather than Frobenius, showing that $r \times s$7 is not uniformly best (Tygert, 2012).

This work suggests one of the clearest templates for a neighbour‑homogeneity score. Replacing the global expectation $r \times s$8 by a local expectation $r \times s$9 defined from a neighborhood $n_{j,k}$0 produces a local Frobenius‑type discrepancy,

$n_{j,k}$1

The source text explicitly proposes neighboring columns homogeneous, neighboring rows homogeneous, and spatial or network neighborhoods as possible constructions, with optional calibration by Monte Carlo under the corresponding local homogeneity model (Tygert, 2012).

3. Graph formulations: label, feature, and edge trustworthiness

In graph learning, neighbour‑homogeneity consistency is formalised much more explicitly. NeuCGC defines a global homophily ratio

$n_{j,k}$2

and a neighborhood homophily ratio

$n_{j,k}$3

The inner term is the node‑level neighborhood homophily of $n_{j,k}$4; the outer average yields a graph‑level summary of how often a typical node’s neighbors share its label. The paper also defines the graph neighborhood congener ratio

$n_{j,k}$5

which measures what fraction of each node’s congeners are actually present in its neighborhood (Peng et al., 17 Dec 2025).

The same paper then introduces a feature‑based unsupervised analogue through the neutral contrastive factor $n_{j,k}$6. Using cross‑view similarity

$n_{j,k}$7

neighbor‑restricted similarity

$n_{j,k}$8

and threshold

$n_{j,k}$9

it defines

$n_{j,}$0

The source characterises $n_{j,}$1 as an unsupervised, feature‑based analogue of $n_{j,}$2, approximating the expected fraction of neighbors that are truly homogeneous in the absence of labels (Peng et al., 17 Dec 2025).

At the edge level, NeuCGC constructs a high‑confidence graph $n_{j,}$3 after fusing embeddings, running K‑means, and selecting top‑$n_{j,}$4 most confident pseudo‑labels. The weight matrix is

$n_{j,}$5

The source text explicitly interprets $n_{j,}$6 as an edge‑level neighbour‑homogeneity consistency score: $n_{j,}$7 means highly consistent or trustworthy, values in $n_{j,}$8 mean moderately consistent, and $n_{j,}$9 means inconsistent or untrustworthy. It further reports that the learned high‑confidence graph $n_{,k}$0 has higher homophily ratio $n_{,k}$1 and much higher congener ratio $n_{,k}$2 than the original graph (Peng et al., 17 Dec 2025).

These quantities yield a natural taxonomy.

Level	Quantity	Interpretation
Node	$n_{,k}$3, $n_{,k}$4	local label or feature homogeneity
Edge	$n_{,k}$5	trustworthiness of a neighbor pair
Graph	$n_{,k}$6, $n_{,k}$7, $n_{,k}$8, $n_{,k}$9	global homophily or neighborhood consistency

The same paper also proposes explicit score forms, including

$n$0

together with graph‑level versions $n$1 and $n$2, and edge‑level $n$3 (Peng et al., 17 Dec 2025).

4. Neighborhood purity and $n$4-hop consistency in graph neural networks

A related but distinct line of work studies neighborhood label purity rather than direct same‑label adjacency. NHGCN defines classical node homophily

$n$5

and observes that this quantity is restricted to $n$6-hop neighbors, depends on the target node’s label, and fails on completely heterophilous graphs such as bipartite graphs where GCN can still perform well. It then introduces Neighborhood Homophily $n$7, defined on a $n$8-hop neighborhood $n$9 as the majority‑label fraction

$(p_1,\dots,p_r)$0

For isolated nodes, the paper sets $(p_1,\dots,p_r)$1. When the neighborhood is nonempty, the range is stated as

$(p_1,\dots,p_r)$2

High values correspond to pure neighborhoods and low label complexity; low values correspond to mixed neighborhoods and high label complexity (Gong et al., 2023).

The conceptual shift is important. Classical homophily asks whether neighbors match the center node’s label. Neighborhood Homophily asks whether the neighborhood itself is label‑consistent. In a bipartite graph where all neighbors have the same but opposite label, $(p_1,\dots,p_r)$3 but $(p_1,\dots,p_r)$4. The paper reports that GCN accuracy rises almost monotonically with NH levels, whereas classical node homophily yields a “tick‑like” relation and does not monotonically explain difficulty. This supports the interpretation of NH as a neighborhood‑level consistency score for message passing (Gong et al., 2023).

A second graph metric, 2‑hop Neighbor Class Similarity, is explicitly aligned with how a simplified 1‑layer GCN propagates label information. For each node $(p_1,\dots,p_r)$5,

$(p_1,\dots,p_r)$6

This is the average, over $(p_1,\dots,p_r)$7’s closed $(p_1,\dots,p_r)$8-hop neighborhood, of the fraction of each neighbor’s closed neighborhood that has the same label as $(p_1,\dots,p_r)$9. It ranges from $$(N_{1,k},\dots,N_{r,k}) \sim \text{Multinomial}\left(n_{,k}; p_1,\dots,p_r\right), \quad k = 1,\dots,s,$$0 to $$(N_{1,k},\dots,N_{r,k}) \sim \text{Multinomial}\left(n_{,k}; p_1,\dots,p_r\right), \quad k = 1,\dots,s,$$1, with $$(N_{1,k},\dots,N_{r,k}) \sim \text{Multinomial}\left(n_{,k}; p_1,\dots,p_r\right), \quad k = 1,\dots,s,$$2 meaning none of the $$(N_{1,k},\dots,N_{r,k}) \sim \text{Multinomial}\left(n_{,k}; p_1,\dots,p_r\right), \quad k = 1,\dots,s,$$3-hop or $$(N_{1,k},\dots,N_{r,k}) \sim \text{Multinomial}\left(n_{,k}; p_1,\dots,p_r\right), \quad k = 1,\dots,s,$$4-hop neighbors of $$(N_{1,k},\dots,N_{r,k}) \sim \text{Multinomial}\left(n_{,k}; p_1,\dots,p_r\right), \quad k = 1,\dots,s,$$5 have label $$(N_{1,k},\dots,N_{r,k}) \sim \text{Multinomial}\left(n_{,k}; p_1,\dots,p_r\right), \quad k = 1,\dots,s,$$6, and $$(N_{1,k},\dots,N_{r,k}) \sim \text{Multinomial}\left(n_{,k}; p_1,\dots,p_r\right), \quad k = 1,\dots,s,$$7 meaning all $$(N_{1,k},\dots,N_{r,k}) \sim \text{Multinomial}\left(n_{,k}; p_1,\dots,p_r\right), \quad k = 1,\dots,s,$$8-hop and $$(N_{1,k},\dots,N_{r,k}) \sim \text{Multinomial}\left(n_{,k}; p_1,\dots,p_r\right), \quad k = 1,\dots,s,$$9-hop neighbors share label $$p_j = \frac{n_{j,}}{n}, \quad e_{j,k} = \frac{n_{j,} \cdot n_{,k}}{n}.$$0. The graph‑level version is the mean over nodes, and a class‑level version averages over nodes in class $$p_j = \frac{n_{j,}}{n}, \quad e_{j,k} = \frac{n_{j,} \cdot n_{,k}}{n}.$$1 (Cavallo et al., 2022).

The paper derives 2NCS from a simplified GCN with $$p_j = \frac{n_{j,}}{n}, \quad e_{j,k} = \frac{n_{j,} \cdot n_{,k}}{n}.$$2 and

$$p_j = \frac{n_{j,}}{n}, \quad e_{j,k} = \frac{n_{j,} \cdot n_{,k}}{n}.$$3

Its argument is that training shapes $$p_j = \frac{n_{j,}}{n}, \quad e_{j,k} = \frac{n_{j,} \cdot n_{,k}}{n}.$$4 by labels in $$p_j = \frac{n_{j,}}{n}, \quad e_{j,k} = \frac{n_{j,} \cdot n_{,k}}{n}.$$5, while inference for $$p_j = \frac{n_{j,}}{n}, \quad e_{j,k} = \frac{n_{j,} \cdot n_{,k}}{n}.$$6 aggregates the $$p_j = \frac{n_{j,}}{n}, \quad e_{j,k} = \frac{n_{j,} \cdot n_{,k}}{n}.$$7 of $$p_j = \frac{n_{j,}}{n}, \quad e_{j,k} = \frac{n_{j,} \cdot n_{,k}}{n}.$$8. The effective label propagation is thus two steps, and 2NCS is designed to measure whether those two steps reinforce the correct label. The source reports that graph‑level accuracy of GCN, GAT, and the simplified GCN increases roughly monotonically with 2NCS, and that node‑level 2NCS shows a stronger monotonic relation with correct classification probability than local homophily or CCNS on Chameleon and Squirrel (Cavallo et al., 2022).

Taken together, NH and 2NCS define two major graph interpretations of neighbour‑homogeneity consistency. NH measures majority purity in a $$p_j = \frac{n_{j,}}{n}, \quad e_{j,k} = \frac{n_{j,} \cdot n_{,k}}{n}.$$9-hop neighborhood. 2NCS measures class‑conditional support transmitted through two hops. The first is target‑label agnostic within the neighborhood; the second is centered on the target node’s label.

5. Nearest neighbors, representativeness, and $\tau$00-consistency

In nearest neighbor matching, the relevant notion of homogeneity is not label purity but substitutability. The problem is to estimate

$\tau$01

where $\tau$02, using a biased sample $\tau$03 with observed outcomes and a missing population $\tau$04 with unobserved outcomes. The method uses Voronoi cells

$\tau$05

the true cell masses $\tau$06, their empirical versions $\tau$07, and the estimator

$\tau$08

The paper proves $\tau$09-consistency of nearest neighbor matching under finite‑dimensional Euclidean covariates and mild integrability conditions, without smoothness or boundedness assumptions on $\tau$10 (Sharpnack, 2019).

The theoretical content can be read as a consistency theory for local homogeneity. One component is nearest‑neighbor regression consistency: $\tau$11 Another is control of the second moment of Voronoi masses; in particular, for $\tau$12-almost all $\tau$13,

$\tau$14

and

$\tau$15

These results formalize when nearest neighbors are homogeneous enough, relative to the target population and the regression function, for local substitution to be asymptotically harmless (Sharpnack, 2019).

The same source proposes data‑driven diagnostics that can be interpreted as neighbour‑homogeneity consistency scores. One is local outcome similarity,

$\tau$16

where $\tau$17 is the nearest neighbor of $\tau$18 among the other observed points. A normalized version is

$\tau$19

A second is Voronoi weight concentration,

$\tau$20

with score

$\tau$21

A third is stability between first and second nearest neighbors,

$\tau$22

The source presents these as heuristic diagnostics tied directly to the proof quantities, not as theorems with finite‑sample guarantees (Sharpnack, 2019).

A plausible implication is that this literature broadens the meaning of neighbour‑homogeneity consistency beyond class agreement. Here, consistency means that local neighborhoods are representative enough to support estimation under biased sampling or missingness.

6. Homogeneity of regression coefficients and CARDS

In high‑dimensional regression, homogeneity means many coefficients are exactly equal and can be grouped. The basic model is

$\tau$23

with a partition $\tau$24 of $\tau$25 such that

$\tau$26

and

$\tau$27

The set $\tau$28 is the zero group, so sparsity becomes a special case of homogeneity with a known atom zero (Ke et al., 2013).

CARDS operationalizes neighbor relations through data‑driven ordering. A preliminary estimator $\tau$29 induces an order

$\tau$30

In this ordered space, neighboring indices are $\tau$31 and $\tau$32. Basic CARDS then solves

$\tau$33

This is a fused, neighbor‑difference penalty promoting piecewise‑constant structure along the ordered sequence (Ke et al., 2013).

Advanced CARDS weakens the need for a fully correct order by introducing ordered segmentation $\tau$34, with segments $\tau$35, and the hybrid penalty

$\tau$36

The first term encourages fusion across adjacent segments; the second encourages within‑segment homogeneity. This creates two flavors of neighborhood: within‑segment neighbors and adjacent‑segment neighbors (Ke et al., 2013).

The paper also contains explicit global consistency measures for group recovery. One is Normalized Mutual Information,

$\tau$37

used in simulations to compare true and estimated partitions. Another is the minimum signal gap relative to noise and group size, which enters the recovery theory through conditions such as

$\tau$38

The paper’s proposed neighbour‑homogeneity score constructions then include average neighbor differences

$\tau$39

penalty‑based scores from the fitted CARDS objective, group‑alignment scores such as NMI, and a gap‑to‑noise index

$\tau$40

These are presented as principled ways to summarize whether neighboring coefficients are consistent with an underlying homogeneous grouping (Ke et al., 2013).

The theoretical gain from exploiting this structure is an oracle rate depending on the number of groups $\tau$41 rather than the ambient dimension $\tau$42: $\tau$43 This rate, together with asymptotic normality and the variance comparison $\tau$44, makes clear that homogeneity consistency is not only descriptive but also statistically consequential (Ke et al., 2013).

7. Interpretation, limitations, and recurrent design principles

Several limitations recur across the literature. In contingency tables, Frobenius discrepancy is often more powerful than classical statistics but not uniformly so, and it lacks the simple $\tau$45 asymptotics available for $\tau$46, $\tau$47, and $\tau$48; Monte Carlo calibration is therefore central (Tygert, 2012). In graph learning, neighborhood trustworthiness depends heavily on the graph’s homophily regime, and low‑homophily or sparsely populated neighborhoods can make local estimates unstable; NeuCGC explicitly addresses this through graph‑level weighting $\tau$49 and edge‑level weighting $\tau$50 (Peng et al., 17 Dec 2025). NHGCN likewise notes dependence on label or pseudo‑label quality, degree sensitivity, and hyperparameters such as $\tau$51 and the threshold $\tau$52 (Gong et al., 2023). In nearest neighbor matching, the theory is confined to finite‑dimensional Euclidean covariates, and the source explicitly warns that the link between nearest‑neighbor diagnostics and asymptotic consistency is not established in general separable metric spaces (Sharpnack, 2019).

A second recurrent issue is that different scores answer different questions. $\tau$53 and $\tau$54 quantify same‑label adjacency. $\tau$55 quantifies neighborhood coverage of the same class. $\tau$56 measures majority purity irrespective of whether the dominant class equals the center node’s class. $\tau$57 measures two‑hop support for the center node’s own label. Frobenius discrepancy measures deviation from a homogeneous expected table. CARDS penalties measure neighborwise equality of parameters. These are all neighbor‑homogeneity consistency measures in a broad sense, but they are not interchangeable (Peng et al., 17 Dec 2025, Gong et al., 2023, Cavallo et al., 2022, Tygert, 2012, Ke et al., 2013).

Despite this heterogeneity, the design pattern is stable. First, define neighborhoods: adjacency sets, $\tau$58-hop neighborhoods, Voronoi cells, ordered neighbors, or local cell neighborhoods. Second, define a homogeneous target: common row proportions, label agreement, majority purity, feature similarity above a threshold, or equal coefficients. Third, quantify consistency through either a proportion, a distance, a divergence, a penalty, or a calibration statistic. Fourth, when asymptotic reference distributions are unavailable or unreliable, calibrate by simulation, pseudo‑labels, or empirical stability analyses (Tygert, 2012, Peng et al., 17 Dec 2025, Sharpnack, 2019).

This suggests that the most precise general characterization of a neighbour‑homogeneity consistency score is methodological rather than taxonomic. It is a score constructed from local structure to quantify how strongly the observed neighborhood supports a homogeneous explanation of the data. Across contingency‑table inference, graph representation learning, nearest‑neighbor estimation, and regression with grouped coefficients, such scores function as diagnostics of model adequacy, guides for weighting or aggregation, and, in some settings, vehicles for measurable gains in statistical power, robustness, or estimation efficiency (Tygert, 2012, Peng et al., 17 Dec 2025, Ke et al., 2013).