Weighted Planted-Partition Models

• Weighted planted-partition models are random graph frameworks where latent community labels and weight mechanisms drive stronger intra-community connections.
• Variants such as degree-corrected SBM, Gaussian weighted SBM, and hypergraph reductions incorporate different weighting schemes to account for node heterogeneity and higher-order interactions.
• These models exhibit sharp recovery and detectability thresholds, highlighting phase transitions in community detection that can be exploited by algorithms like spectral methods and semidefinite relaxation.

The weighted planted-partition model denotes a family of random graph and hypergraph models in which a latent partition determines stronger within-community than between-community affinity, while “weighting” is introduced in different ways across the literature: as observed edge weights, as latent node-specific propensities that modulate connection probabilities, or as pairwise weights induced from higher-order interactions. In the Bernoulli special case, it reduces to the classical planted partition model, also known as the stochastic block model (SBM), with within-block and between-block probabilities determined solely by community membership (Mossel et al., 2012, Gulikers et al., 2015, Pandey et al., 2024, Ghoshdastidar et al., 2015).

1. Definitions and principal variants

In its classical sparse form, the planted partition model assigns each vertex a latent label and places each edge independently with probability $a/n$ within communities and $b/n$ across communities. This is the unweighted Bernoulli baseline from which several weighted generalizations depart. The main distinctions concern what is being weighted: the observed edge, the latent vertex propensity, or a pairwise reduction of higher-order structure (Mossel et al., 2012, Gulikers et al., 2015, Ghoshdastidar et al., 2015).

Variant	Primitive random object	Weight interpretation
Sparse SBM / planted partition	Bernoulli edge indicator	Binary $\{0,1\}$ special case
Degree-corrected SBM	Bernoulli edge with factor $\phi_u\phi_v$	Latent node propensities
Gaussian weighted SBM	Real-valued $A_{ij}$	Observed edge weights
Hypergraph planted partition with pairwise reduction	Bernoulli hyperedges	Induced pairwise weights via $H\Delta^{-1}H^T$

This taxonomy matters because superficially similar models can have different statistical structure. The degree-corrected model is “weighted” through heterogeneous expected degrees rather than through observed weighted edges, whereas the Gaussian weighted SBM is genuinely an edge-weight model. The hypergraph setting is different again: the original object is an unweighted random hypergraph, but the analysis proceeds through a weighted graph reduction with block-structured expectation (Gulikers et al., 2015, Pandey et al., 2024, Ghoshdastidar et al., 2015).

2. Degree correction as a weighted planted-partition mechanism

A canonical sparse weighted planted-partition model is the degree-corrected stochastic block model. Here each vertex $u$ receives a latent class label $\sigma_u \in \{1,\dots,q\}$ and an i.i.d. weight $\phi_u$ drawn from a law $\nu$ on $b/n$0, with moments $b/n$1. Conditioned on labels and weights, distinct vertices $b/n$2 are connected independently with probabilities

$b/n$3

If $b/n$4, the model reduces to the ordinary sparse planted-partition model. The expected degree of a vertex with weight $b/n$5 satisfies

$b/n$6

so the latent weight scales expected degree linearly. The model allows heavy-tailed weights, assuming a finite second moment and, for some technical arguments, a power-law tail upper bound with exponent $b/n$7 (Gulikers et al., 2015).

Its main information-theoretic result is a non-reconstruction theorem. If

$b/n$8

then no estimator can recover the communities with positive correlation from a single observed graph when the weights are unknown. The paper formalizes positive correlation by requiring the existence of $b/n$9 such that

$\{0,1\}$0

A stronger statement is proved first: for uniformly random vertices $\{0,1\}$1 and any $\{0,1\}$2,

$\{0,1\}$3

Thus even the whole graph and the label of another random vertex become asymptotically useless for predicting the label of a typical vertex below threshold (Gulikers et al., 2015).

The local structure is controlled by a multi-type branching process rather than a homogeneous Galton–Watson tree. The mean offspring number is

$\{0,1\}$4

and the broadcast noise parameter is

$\{0,1\}$5

The appearance of $\{0,1\}$6, rather than $\{0,1\}$7, reflects size biasing in local neighborhoods: high-weight vertices are disproportionately sampled during exploration. In this sense, the second moment of the latent weight law is the parameter through which degree heterogeneity modifies detectability. The same work also establishes a precise local coupling for radius-$\{0,1\}$8 neighborhoods and notes, as a by-product, that long-range interactions are weak in power-law degree-corrected SBMs with sufficiently large exponent (Gulikers et al., 2015).

3. Observed edge weights: the Gaussian weighted SBM

A distinct weighted planted-partition model replaces Bernoulli edges by real-valued observations. In the Gaussian weighted stochastic block model with two symmetric communities, the unknown label vector is balanced,

$\{0,1\}$9

and for each distinct pair $\phi_u\phi_v$0,

$\phi_u\phi_v$1

The assortative regime is $\phi_u\phi_v$2. Under the exact-recovery scaling,

$\phi_u\phi_v$3

with signal-to-noise ratio

$\phi_u\phi_v$4

This model is the Gaussian analogue of the balanced two-community planted bisection/SBM (Pandey et al., 2024).

The exact-recovery threshold is sharp: $\phi_u\phi_v$5 If $\phi_u\phi_v$6, no estimator can exactly recover the community structure with probability bounded away from zero; if $\phi_u\phi_v$7, the maximum likelihood estimator succeeds with probability approaching one. The MLE reduces to the balanced quadratic optimization

$\phi_u\phi_v$8

where $\phi_u\phi_v$9. The same threshold is achieved algorithmically by both a semidefinite relaxation,

$A_{ij}$0

and a spectral estimator based on the top eigenvector of a centered matrix $A_{ij}$1. The paper therefore establishes, for exact recovery in this model, an absence of an information-computation gap (Pandey et al., 2024).

The same work compares this symmetric weighted planted-partition problem with recovering a planted dense weighted subgraph. For a planted community of size $A_{ij}$2, statistical impossibility holds when $A_{ij}$3, while MLE and SDP succeed when $A_{ij}$4. Accordingly, exact recovery of two symmetric communities is strictly easier than exact recovery of a single planted community of size $A_{ij}$5 in the Gaussian weighted setting (Pandey et al., 2024).

4. Reconstruction, detectability, and threshold principles

The sparse unweighted planted partition model already exhibits the central threshold phenomenon. For the two-community model $A_{ij}$6, impossibility of nontrivial reconstruction holds when

$A_{ij}$7

and, under the strict inequality, the model is mutually contiguous with the Erdős–Rényi graph $A_{ij}$8. In the same regime one cannot even consistently estimate $A_{ij}$9 and $H\Delta^{-1}H^T$0. Above threshold,

$H\Delta^{-1}H^T$1

the parameters can be estimated efficiently from short-cycle counts. The conceptual explanation is a local weak limit to a broadcast process on a Galton–Watson tree with offspring mean $H\Delta^{-1}H^T$2 and channel parameter $H\Delta^{-1}H^T$3, so the criterion becomes

$H\Delta^{-1}H^T$4

This is the Kesten–Stigum form for the sparse Bernoulli planted partition model (Mossel et al., 2012).

The degree-corrected threshold is a direct generalization of that principle. The multi-type branching process generated by latent weights has effective branching factor proportional to $H\Delta^{-1}H^T$5, and the non-reconstruction criterion becomes

$H\Delta^{-1}H^T$6

leading to the impossibility theorem at

$H\Delta^{-1}H^T$7

For $H\Delta^{-1}H^T$8, the conditional mean adjacency matrix has “mean-eigenvalues”

$H\Delta^{-1}H^T$9

so the threshold can be written as

$u$0

This exhibits degree heterogeneity as a rescaling of signal and noise through the second weight moment (Gulikers et al., 2015).

These thresholds concern correlated reconstruction or non-reconstruction in sparse Bernoulli or degree-corrected graphs. By contrast, the Gaussian weighted SBM addresses exact recovery under a different observation model and scaling, with means of order $u$1 and threshold $u$2. The shared feature is not a single universal formula, but rather the emergence of a sharp boundary between statistically feasible and infeasible recovery regimes (Pandey et al., 2024).

5. Hypergraph reductions and local-support algorithms

One route to weighted planted-partition structure begins with higher-order interactions. In the planted partition model for sparse random non-uniform hypergraphs, an $u$3-edge $u$4 is included independently with probability

$u$5

for $u$6. The associated weighted pairwise matrix is

$u$7

and the normalized Laplacian is

$u$8

Its population expectation has block form

$u$9

which is analogous to the block expectation of a weighted SBM, although the weights are induced by hyperedge-to-pairwise reduction rather than by primitive pairwise weighted edges. The paper proves consistency of spectral hypergraph partitioning under this model and explicitly notes that the framework can be extended to weighted hypergraphs when $\sigma_u \in \{1,\dots,q\}$0 and the first moment matches the Bernoulli model (Ghoshdastidar et al., 2015).

A different adjacent line of work exploits second-order local structure. In streaming partitioning for the planted partition model $\sigma_u \in \{1,\dots,q\}$1, class assignment based on length-2 walks or common neighbors yields expected same-cluster versus different-cluster separation

$\sigma_u \in \{1,\dots,q\}$2

which explains why path-2 information can succeed even when $\sigma_u \in \{1,\dots,q\}$3. Among constant walk lengths $\sigma_u \in \{1,\dots,q\}$4, the analysis identifies $\sigma_u \in \{1,\dots,q\}$5 as optimal in the sense of requiring the weakest asymptotic lower bound on $\sigma_u \in \{1,\dots,q\}$6 (Tsourakakis, 2014).

For many-community regimes with highly variable block sizes, Diamond Percolation retains only edges participating in at least two triangles, then outputs the connected components of the retained graph. In the unweighted planted partition model it yields exact, almost exact, or weak recovery under conditions expressed through $\sigma_u \in \{1,\dots,q\}$7, $\sigma_u \in \{1,\dots,q\}$8, and the community-size distribution, including power-law partitions. The paper introduces a correlation coefficient

$\sigma_u \in \{1,\dots,q\}$9

to compare partitions when the number and sizes of communities vary. These common-neighbor and triangle-based results are not weighted theorems, but they indicate that second-order local support can remain informative when direct within-versus-between separation is weak. A plausible implication is that weighted analogues should be built from weighted 2-walk or weighted triangle scores rather than from first-order edge statistics alone (Gösgens et al., 2 Apr 2025).

6. Objectives, metrics, and recurrent misconceptions

A recurrent source of confusion is that “weighted planted partition” does not refer to a single canonical construction. In the degree-corrected model, the weights are latent node propensities $\phi_u$0 and there are no observed weighted edges. In the hypergraph model, the original random object is an unweighted hypergraph, and weights enter only after projection to the pairwise matrix $\phi_u$1. Only the Gaussian weighted SBM is a direct pairwise edge-weight model in which the observed $\phi_u$2 are real-valued random variables with class-dependent distributions (Gulikers et al., 2015, Ghoshdastidar et al., 2015, Pandey et al., 2024).

The recovery objective also changes across models. In the degree-corrected sparse setting, the natural benchmark is positive correlation with the true labels; in the Gaussian weighted SBM, the benchmark is exact recovery up to global sign; and in the many-community planted partition setting with highly unequal sizes, a correlation coefficient between partitions becomes preferable to label-alignment metrics (Gulikers et al., 2015, Pandey et al., 2024, Gösgens et al., 2 Apr 2025).

Optimization objectives need not coincide with inference objectives. In the sparse planted bisection model with $\phi_u$3, the minimum bisection width satisfies an asymptotic formula derived from a fixed point of a message-distribution operator on a two-type Galton–Watson tree, under the condition

$\phi_u$4

Yet the planted bisection is generally not the minimum bisection in this regime and may differ from the optimum by $\phi_u$5. This shows that recovering planted communities and minimizing a balanced cut are distinct problems even in the binary-weight special case (Coja-Oghlan et al., 2015).

Taken together, these results identify the weighted planted-partition model not as one fixed object but as a class of latent-partition models in which block structure is preserved while weights encode degree heterogeneity, observed affinity strength, or higher-order interaction reduction. The central technical themes are the same across these variants: block-structured expectations, local weak limits, sharp threshold phenomena, and the need to distinguish carefully between statistical recovery, parameter estimation, and combinatorial optimization (Mossel et al., 2012, Gulikers et al., 2015, Pandey et al., 2024).