Disentangled Multimodal Graph Clustering (DMGC)

• The paper introduces DMGC, an unsupervised framework that disentangles homophilic and heterophilic structural signals in multimodal graphs.
• It constructs separate views—a homophily-enhanced graph and modality-specific heterophily-aware graphs—and fuses them through dual-frequency filtering.
• Self-supervised learning aligns modality-specific embeddings with DEC-style clustering, achieving state-of-the-art results on diverse graph datasets.

Disentangled Multimodal Graph Clustering (DMGC) is an unsupervised clustering framework for multimodal graphs that explicitly separates homophilic and heterophilic structural signals instead of treating graph propagation as governed by a single neighborhood regime. Introduced in "Disentangling Homophily and Heterophily in Multimodal Graph Clustering" (Guo et al., 21 Jul 2025), the method is built for multimodal graphs that combine unstructured heterogeneous data with structured interconnections, and it is motivated by the observation that real-world multimodal graphs often exhibit hybrid neighborhood patterns rather than being purely homophilic or purely heterophilic. DMGC decomposes the original graph into a homophily-enhanced graph and modality-specific heterophily-aware graphs, processes them through a Multimodal Dual-frequency Fusion mechanism, and trains the resulting representation with self-supervised alignment and clustering objectives. Its reported empirical profile is state-of-the-art across both multimodal and multi-relational graph datasets.

1. Problem setting and hybrid neighborhood structure

DMGC formulates a multimodal graph as

$$\mathcal{G} = (\mathcal{V}, \mathcal{A}, \mathcal{X}),$$

where $\mathcal{V}$ is the node set with $|\mathcal{V}| = N$, $\mathcal{A} = \{A_1, A_2, \dots, A_R\}$ is a set of $R$ adjacency matrices representing different relations, and $\mathcal{X} = \{X^1, X^2, \dots, X^M\}$ is the set of multimodal node features, one matrix per modality. Each $X^i \in \mathbb{R}^{N \times d_f}$ is a modality-specific feature matrix projected into a shared dimension using pretrained encoders such as CLIP. The task is to discover latent classes without labels by using both topology and multimodal content (Guo et al., 21 Jul 2025).

The central empirical observation is that node-level homophily ratios in real multimodal graphs are broadly distributed and often moderate rather than extreme. A node neighborhood therefore often contains both homophilic neighbors, which are class-consistent, and heterophilic neighbors, which are class-divergent. The paper characterizes this as a hybrid neighborhood pattern. In this regime, standard message passing can aggregate both helpful and harmful signals: in moderately homophilic graphs it can blur class boundaries, while in heterophilic regions it can propagate misleading features and increase category confusion.

DMGC treats homophily and heterophily as complementary rather than mutually exclusive. Homophily is used to discover cross-modal class consistency, whereas heterophily is treated as preserving modality-specific inter-class distinctions. The method’s foundational claim is therefore not that one structural regime should dominate, but that both should be separated and exploited jointly. This establishes a notion of disentanglement that is graph-structural and frequency-aware rather than merely latent-factor factorization.

2. Disentangled graph construction

The first major component is Disentangled Graph Construction, which builds separate graph views from the original hybrid graph. For each modality $i$ and relation $r$, raw features are first propagated over the original relation graph: $$X^{i(0)}_{r} = X^i,\quad X^{i(t+1)}_{r} = (1 - \alpha)\tilde{A}_r X^{i(t)}_{r} + \alpha X^i,$$ where $\mathcal{V}$0 is the normalized adjacency for relation $\mathcal{V}$1, $\mathcal{V}$2, and $\mathcal{V}$3 is a residual coefficient. After $\mathcal{V}$4 steps,

$\mathcal{V}$5

The aggregated features across relations are then concatenated: $\mathcal{V}$6

From these relation-aware modality features, DMGC constructs two complementary structural objects. The first is a homophily-enhanced graph. A cross-modality consensus feature is defined as

$\mathcal{V}$7

followed by the similarity matrix

$\mathcal{V}$8

The homophily-enhanced graph is formed by selecting the top-$\mathcal{V}$9 nearest neighbors,

$|\mathcal{V}| = N$0

and normalizing the result as

$|\mathcal{V}| = N$1

This graph is designed to capture strong class-consistent cross-modal agreement.

The second object is a family of heterophily-aware graphs, one per modality. For each modality $|\mathcal{V}| = N$2, the method constructs an intra-modality similarity matrix

$|\mathcal{V}| = N$3

then uses the complement $|\mathcal{V}| = N$4 to identify dissimilar neighbors: $|\mathcal{V}| = N$5 The corresponding operator is expressed in Laplacian form as

$|\mathcal{V}| = N$6

The homophily graph therefore emphasizes similarity and cross-modal consensus, whereas the heterophily graphs explicitly emphasize dissimilarity and modality-specific inter-class structure.

The paper also addresses construction cost. Because exact similarity computation can be quadratic, DMGC proposes locality-sensitive hashing for the homophilic graph and anchor-based construction for heterophilic graphs. The stated complexity is

$|\mathcal{V}| = N$7

which is described as linear in graph size terms and scalable to larger datasets such as Ele-Fashion.

3. Multimodal Dual-frequency Fusion

The second major component is Multimodal Dual-frequency Fusion, which is the method’s core representational mechanism. Each aggregated feature is first encoded through a relation-specific projection: $|\mathcal{V}| = N$8 where $|\mathcal{V}| = N$9 is a learnable encoder.

DMGC then applies two complementary graph filters. A low-pass pathway operates on the homophily-enhanced graph $\mathcal{A} = \{A_1, A_2, \dots, A_R\}$0, and a high-pass pathway operates on the heterophilic Laplacian $\mathcal{A} = \{A_1, A_2, \dots, A_R\}$1. With initialization

$\mathcal{A} = \{A_1, A_2, \dots, A_R\}$2

the iterative updates are

$\mathcal{A} = \{A_1, A_2, \dots, A_R\}$3

$\mathcal{A} = \{A_1, A_2, \dots, A_R\}$4

After $\mathcal{A} = \{A_1, A_2, \dots, A_R\}$5 layers, the outputs are averaged across relations: $\mathcal{A} = \{A_1, A_2, \dots, A_R\}$6

These two branches are then combined by an adaptive gate: $\mathcal{A} = \{A_1, A_2, \dots, A_R\}$7 where $\mathcal{A} = \{A_1, A_2, \dots, A_R\}$8 is trainable. The paper’s interpretation is explicit: low-pass filtering smooths and strengthens class-consistent semantics, high-pass filtering preserves distinctions and outlier-like structure, and the gate prevents overcommitment to either extreme. In the language used by the method, the final representation is built from both low-frequency, class-smoothing information and high-frequency, class-separating information.

A final multimodal fusion stage aggregates modality-specific embeddings by attention. The attention scores are

$\mathcal{A} = \{A_1, A_2, \dots, A_R\}$9

with normalized weights

$R$0

and the unified embedding

$R$1

This makes the joint representation modality-aware: more informative modalities receive larger fusion weights.

4. Self-supervised learning and clustering objective

DMGC is trained without labels. Its objective combines reconstruction, alignment, and clustering terms: $R$2 Each term corresponds to a distinct self-supervised constraint (Guo et al., 21 Jul 2025).

The intra-modality reconstruction term preserves modality-specific information. The low-pass and high-pass embeddings are concatenated and decoded: $R$3 A cosine reconstruction loss then compares $R$4 with the original feature $R$5: $R$6 This is intended to prevent the two frequency channels from collapsing feature information.

The dual-frequency alignment term aligns each modality’s low- and high-frequency representations with its fused representation: $R$7 where $R$8 is an InfoNCE-style node-level contrastive loss. This keeps both frequency channels useful while allowing them to remain distinct.

The cross-modality alignment term enforces semantic agreement across modalities: $R$9 using the pairwise contrastive form

$\mathcal{X} = \{X^1, X^2, \dots, X^M\}$0

where $\mathcal{X} = \{X^1, X^2, \dots, X^M\}$1 is cosine similarity and $\mathcal{X} = \{X^1, X^2, \dots, X^M\}$2 is temperature.

The clustering objective is DEC-style. Soft cluster assignment is defined by Student’s $\mathcal{X} = \{X^1, X^2, \dots, X^M\}$3-distribution: $\mathcal{X} = \{X^1, X^2, \dots, X^M\}$4 with a sharpened target distribution

$\mathcal{X} = \{X^1, X^2, \dots, X^M\}$5

The resulting clustering loss is

$\mathcal{X} = \{X^1, X^2, \dots, X^M\}$6

Unlike approaches that only learn embeddings and then apply post-hoc clustering, DMGC integrates cluster refinement into training.

5. Empirical evaluation

The reported experiments span six datasets and use ACC, NMI, and ARI as metrics (Guo et al., 21 Jul 2025). The datasets include multimodal multi-relational graphs, multi-relational graphs, and a large multimodal graph: IMDB, Amazon, ACM, DBLP, Yelp, and Ele-Fashion. Additional tests are reported on Texas and Chameleon as heterophilic graphs, and robustness is evaluated under 20% Gaussian noise on Amazon.

Dataset	Setting	Reported ACC / NMI / ARI
IMDB	text + image; 3 classes; 4278 nodes	58.16 / 14.01 / 17.32
Amazon	text + image; 3 classes; 6158 nodes	79.83 / 37.13 / 46.28
Ele-Fashion	text + image; 11 classes; 97,766 nodes	51.17 / 51.71 / 49.02
ACM	text; 3 classes; 4019 nodes	91.57 / 70.38 / 75.46
DBLP	text; 4 classes; 2957 nodes	91.14 / 73.76 / 78.79
Yelp	text; 3 classes; 2614 nodes	92.77 / 73.63 / 77.35

The baseline set is divided by problem type. For multimodal graph clustering, the paper compares against KMeans on text/image features, HDMI, InfoMGF-RA, GWMAC, GCFAgg, and MHGAT adapted to unsupervised setting. For multi-relational graph clustering, it compares against VGAE, DGI, O2MAC, MvAGC, MCGC, HDMI, MGDCR, DMG, BTGF, and InfoMGF-RA. The main reported pattern is that methods ignoring graph topology underperform, methods ignoring rich multimodal information also underperform, and the hybrid structural regime requires a method that models both.

Ablation studies remove $\mathcal{X} = \{X^1, X^2, \dots, X^M\}$7, $\mathcal{X} = \{X^1, X^2, \dots, X^M\}$8, $\mathcal{X} = \{X^1, X^2, \dots, X^M\}$9, the homophily graph, and the heterophily graph. Removing any of these components hurts performance. The paper further reports that homophily is more dominant, but heterophily still adds meaningful complementary information, and that both graph views are necessary. Additional analysis shows that DMGC improves the homophily ratio of constructed graphs and produces cleaner t-SNE clusters with higher silhouette scores.

These results support the paper’s core claim that clustering performance benefits from an explicit decomposition of hybrid structure. A plausible implication is that the model’s gain comes less from any single module in isolation than from the coordination of graph construction, dual-frequency propagation, and self-supervised alignment.

6. Relation to adjacent disentangled graph learning

Within the provided literature, DMGC occupies a distinct position relative to earlier disentangled graph models. "Independence Promoted Graph Disentangled Networks" (Liu et al., 2019) addresses disentangled representation learning with independent latent factors in graph convolutional networks. It replaces holistic neighborhood aggregation with neighborhood routing, splits node representations into $X^i \in \mathbb{R}^{N \times d_f}$0 channels, and uses the Hilbert-Schmidt Independence Criterion as a regularizer at the output layer to enforce independence among latent representations. For clustering, the learned node embeddings are extracted and $X^i \in \mathbb{R}^{N \times d_f}$1-means is run with the number of clusters set equal to the number of classes. IPGDN is therefore a disentangled graph convolutional framework whose clustering value lies in the geometry of single-graph embeddings rather than in multimodal fusion or an integrated clustering objective.

"Graph-based Unsupervised Disentangled Representation Learning via Multimodal LLMs" (Xie et al., 2024) is closer to DMGC in its use of multimodal information and graph-structured inductive bias, but it is not a graph clustering method in the usual sense. That work proposes GEM, which combines a $X^i \in \mathbb{R}^{N \times d_f}$2-VAE branch, an MLLM branch using GPT-4o as a relation predictor, and a bidirectional weighted graph whose nodes are latent factors and whose edges encode weighted directed correlations. Its graph models factor interdependence rather than sample-to-sample similarity, and it does not present a clustering objective, cluster assignment module, or graph-level benchmark in the style of DMGC.

These comparisons clarify an important conceptual distinction. In DMGC, disentanglement refers to separating homophilic and heterophilic structural signals, then processing them through low-pass and high-pass pathways for unsupervised clustering. In IPGDN, disentanglement refers to multiple latent components per node with explicit independence promotion. In GEM, disentanglement is relation-aware factor discovery that explicitly rejects independence as a universal assumption. The three methods therefore share a broad concern with non-holistic graph representation, but they operationalize “disentanglement” at different levels: graph structure and frequency in DMGC, independent latent channels in IPGDN, and correlated latent factors in GEM.

7. Scope, interpretive issues, and limitations

Several interpretive issues recur in discussions of DMGC. One is the assumption that heterophily should be treated as noise in clustering. The method explicitly rejects that view: heterophily is used to preserve modality-specific inter-class distinctions and is modeled through dedicated heterophily-aware graphs rather than suppressed. Another is the assumption that multimodal graph clustering can be handled by simply extending a homophily-oriented GNN. The paper argues that this is inadequate when node-level homophily ratios are moderate and broadly distributed, because standard propagation can mix class-consistent and class-divergent signals too early.

A further issue concerns the meaning of disentanglement. DMGC does not frame disentanglement primarily as statistical independence among latent factors. Instead, it disentangles complementary graph views and complementary frequency components. This distinguishes it from independence-promoted graph disentanglement and from latent-factor graphs that explicitly model correlations.

The paper also notes practical constraints. Graph construction can be expensive without scalable approximations, and performance depends on hyperparameters such as $X^i \in \mathbb{R}^{N \times d_f}$3, $X^i \in \mathbb{R}^{N \times d_f}$4, $X^i \in \mathbb{R}^{N \times d_f}$5, and $X^i \in \mathbb{R}^{N \times d_f}$6. Overly large values can degrade quality by mixing frequency signals too much or over-aligning modalities. These are implementation-level limitations rather than conceptual failures, but they delimit the method’s operating regime.

Taken together, the reported evidence presents DMGC as a framework for multimodal graph clustering under hybrid neighborhood patterns: it builds a homophily-enhanced graph, constructs modality-specific heterophily-aware graphs, fuses them through dual-pass frequency filtering, and learns cluster-friendly embeddings through self-supervision and DEC-style refinement. Its main significance lies in making hybrid structure a first-class modeling object rather than a nuisance to be averaged away.