Tri-Learn Graph Fusion Network

• The paper’s main contribution is a tri-branch architecture that fuses autoencoder, GCN, and Transformer methods to overcome over-smoothing and over-compression in graph clustering.
• It integrates local topological aggregation, attribute reconstruction, and long-range dependency modeling through layerwise tri-learning and a triple-channel enhancement module.
• Empirical results demonstrate significant performance improvements on datasets such as ACM, Reuters, and USPS, validated by dual self-supervised clustering and reconstruction objectives.

Searching arXiv for the Tri-GFN paper and closely related graph-fusion work to ground the article in current preprints. Tri-Learn Graph Fusion Network (Tri-GFN) is a deep framework for attributed graph clustering that combines a Graph Convolutional Network (GCN), an Autoencoder (AE), and a Graph Transformer in a single tri-branch architecture. Its stated purpose is to address over-smoothing and over-compression in graph clustering by coupling local topological aggregation, attribute-preserving representation learning, and long-range dependency modeling through a tri-learning mechanism, a triple-channel enhancement module, and dual self-supervised clustering objectives (Li et al., 18 Jul 2025).

1. Model identity, task setting, and naming

Tri-GFN is formulated for attributed graph clustering on an input graph

$G=(V,E,X,A),$

where $V=\{v_1,\dots,v_N\}$ is the node set, $E$ is the edge set, $X\in\mathbb{R}^{N\times d_0}$ is the node-attribute matrix, and $A\in\mathbb{R}^{N\times N}$ is the adjacency matrix (Li et al., 18 Jul 2025). The task is unsupervised: nodes are partitioned into $k$ clusters without labels, using both attributes and topology.

The architecture uses both dense and sparse graph encodings. The normalized adjacency is

$\tilde A = D^{-\frac12}(A+I)D^{-\frac12},$

with

$D_{ii}=\sum_j (A+I)_{ij},$

while the sparse edge-index encoding is

$E=[s;t]\in\mathbb{R}^{2\times \delta},$

where $\delta$ is the number of edges. A notable implementation detail is that the manuscript uses both $V=\{v_1,\dots,v_N\}$0 and $V=\{v_1,\dots,v_N\}$1 somewhat inconsistently: $V=\{v_1,\dots,v_N\}$2 appears in the GCN and Graph Transformer propagation formulas, whereas $V=\{v_1,\dots,v_N\}$3 and $V=\{v_1,\dots,v_N\}$4 are used in enhancement and reconstruction.

A recurrent source of confusion is the similarity between the acronym Tri-GFN and the earlier Question-Driven Graph Fusion Network, QD-GFN, for visual question answering. QD-GFN is a distinct model: it is not called “Tri-Learn Graph Fusion Network” or “Tri-GFN,” even though it is conceptually a three-graph fusion architecture because it fuses semantic, spatial, and implicit relation graphs (Qian et al., 2022). In contrast, Tri-GFN denotes the specific attributed graph clustering framework built from GCN, AE, and Graph Transformer (Li et al., 18 Jul 2025).

2. Architectural composition and branch roles

Tri-GFN is organized as three parallel but interacting branches. The AE branch encodes node attributes and reconstructs the original feature matrix; the GCN branch captures local structural patterns through neighborhood aggregation; the Graph Transformer branch models attention-weighted, non-local dependencies (Li et al., 18 Jul 2025).

The AE encoder is defined as

$V=\{v_1,\dots,v_N\}$5

and the decoder is

$V=\{v_1,\dots,v_N\}$6

The feature reconstruction loss is

$V=\{v_1,\dots,v_N\}$7

The final AE embedding is denoted $V=\{v_1,\dots,v_N\}$8.

The GCN branch is written in compressed form as

$V=\{v_1,\dots,v_N\}$9

with the node-wise propagation rule

$E$0

This branch is responsible for local topology encoding and graph reconstruction.

The Graph Transformer branch is defined analogously by

$E$1

Its attention mechanism uses the projections

$E$2

and the attention coefficient

$E$3

The resulting aggregation is

$E$4

The manuscript is internally inconsistent about normalization in this branch: Eq. (8) uses a nonlinear $E$5 described as sigmoid, whereas the architecture discussion also mentions softmax. This suggests an ambiguity in exposition rather than in the high-level design.

3. Tri-learning mechanism and triple-channel enhancement

The distinctive feature of Tri-GFN is not merely the coexistence of three encoders, but their explicit coupling through layerwise interaction and final fusion (Li et al., 18 Jul 2025).

At each layer, the AE representation is injected into both the GCN and Graph Transformer streams. For the GCN branch,

$E$6

or, equivalently,

$E$7

For the Graph Transformer branch,

$E$8

or

$E$9

Here $X\in\mathbb{R}^{N\times d_0}$0 controls the balance between AE-guided information and branch-specific hidden state. This layerwise interpolation is the paper’s clearest operational definition of “tri-learning.”

The final triple-channel enhancement module fuses the three terminal branch outputs by weighted summation followed by graph diffusion:

$X\in\mathbb{R}^{N\times d_0}$1

The manuscript’s OCR is inconsistent around the symbols for these coefficients, but the intended semantics are clear: $X\in\mathbb{R}^{N\times d_0}$2, $X\in\mathbb{R}^{N\times d_0}$3, and $X\in\mathbb{R}^{N\times d_0}$4 are scalar weights for the GCN, AE, and Graph Transformer channels, respectively, and the parameter analysis searches combinations satisfying

$X\in\mathbb{R}^{N\times d_0}$5

This suggests a two-stage enhancement logic. First, AE features anchor the structural branches at every depth; second, the final representations are merged and re-propagated with $X\in\mathbb{R}^{N\times d_0}$6. A plausible implication is that Tri-GFN uses the AE branch as a stabilizing attribute prior while still allowing the GCN and Transformer branches to encode complementary structural information.

4. Clustering formulation and optimization objective

Tri-GFN adopts a DEC-style self-training procedure for clustering (Li et al., 18 Jul 2025). Given node embedding $X\in\mathbb{R}^{N\times d_0}$7 and centroid $X\in\mathbb{R}^{N\times d_0}$8, the soft assignment is

$X\in\mathbb{R}^{N\times d_0}$9

The target distribution is

$A\in\mathbb{R}^{N\times N}$0

and the clustering loss is

$A\in\mathbb{R}^{N\times N}$1

A second assignment distribution $A\in\mathbb{R}^{N\times N}$2 is produced from the AE representation, and a consistency term aligns the fused and AE-side clusterings:

$A\in\mathbb{R}^{N\times N}$3

The paper does not explicitly print the full equation used to compute $A\in\mathbb{R}^{N\times N}$4, but it states that $A\in\mathbb{R}^{N\times N}$5 is generated from the AE representation and that the final prediction uses it:

$A\in\mathbb{R}^{N\times N}$6

That choice is unusual: the fused representation generates $A\in\mathbb{R}^{N\times N}$7, but the reported final cluster label is derived from the AE-side distribution $A\in\mathbb{R}^{N\times N}$8.

Reconstruction is also part of the objective. The paper defines

$A\in\mathbb{R}^{N\times N}$9

where $k$0 is a feature-matrix reconstruction term, $k$1 is an adjacency reconstruction term, and

$k$2

It also introduces an additional AE-related reconstruction term $k$3. The printed formula around the intermediate variable $k$4 is corrupted, so only the structure of the loss is secure. The total objective is stated clearly:

$k$5

Optimization proceeds in two phases. The AE is pretrained for 50 epochs with learning rate $k$6. K-means is then run on the pretrained AE embedding, with 20 random restarts, to initialize the cluster centroids. Joint training subsequently updates all branches and losses together with Adam (Li et al., 18 Jul 2025).

5. Empirical performance, datasets, and ablations

Tri-GFN is evaluated on seven public datasets: ACM, DBLP, Citeseer, HHAR, Cora, USPS, and Reuters (Li et al., 18 Jul 2025). The reported metrics are ACC, NMI, ARI, and Macro F1-score. Baselines include KMEANS, AE, DEC, IDEC, GAE, VGAE, DAEGC, ARGA, DDGAE, DCRN, TDCN, SDCN, and MBN.

The strongest headline results concern ACM, Reuters, and USPS. On ACM, Tri-GFN reports $k$7 ACC, $k$8 NMI, approximately $k$9 ARI, and $\tilde A = D^{-\frac12}(A+I)D^{-\frac12},$0 F1, corresponding to about $\tilde A = D^{-\frac12}(A+I)D^{-\frac12},$1 ACC improvement over MBN. On Reuters, it reports $\tilde A = D^{-\frac12}(A+I)D^{-\frac12},$2 ACC, $\tilde A = D^{-\frac12}(A+I)D^{-\frac12},$3 NMI, $\tilde A = D^{-\frac12}(A+I)D^{-\frac12},$4 ARI, and $\tilde A = D^{-\frac12}(A+I)D^{-\frac12},$5 F1; the paper highlights an ACC improvement of $\tilde A = D^{-\frac12}(A+I)D^{-\frac12},$6 over MBN. On USPS, it reports $\tilde A = D^{-\frac12}(A+I)D^{-\frac12},$7 ACC, $\tilde A = D^{-\frac12}(A+I)D^{-\frac12},$8 NMI, $\tilde A = D^{-\frac12}(A+I)D^{-\frac12},$9 ARI, and $D_{ii}=\sum_j (A+I)_{ij},$0 F1, with a highlighted $D_{ii}=\sum_j (A+I)_{ij},$1 ACC gain over MBN (Li et al., 18 Jul 2025).

Dataset	Tri-GFN ACC	Highlighted ACC gain
ACM	$D_{ii}=\sum_j (A+I)_{ij},$2	$D_{ii}=\sum_j (A+I)_{ij},$3 over MBN
Reuters	$D_{ii}=\sum_j (A+I)_{ij},$4	$D_{ii}=\sum_j (A+I)_{ij},$5 over MBN
USPS	$D_{ii}=\sum_j (A+I)_{ij},$6	$D_{ii}=\sum_j (A+I)_{ij},$7 over MBN

The implementation details reported in the manuscript include AE, GCN, and Transformer layer dimensions $D_{ii}=\sum_j (A+I)_{ij},$8; Adam optimization; fixed random seed; batch size 256; and a hardware/software environment consisting of an Intel i5-1240P CPU, NVIDIA GTX 3070 GPU, 32GB RAM, Windows 10, and “PyTorch 3.9.0” (Li et al., 18 Jul 2025).

Ablation studies support the tri-branch formulation. Removing the AE branch causes the largest degradation overall, removing the GCN branch significantly reduces NMI and ARI, and removing the Transformer branch produces smaller but consistent drops. Decoder removal also degrades results, indicating that reconstruction remains functional rather than decorative. The depth study reports Tri-GFN-3 as the best overall setting, while four propagation layers can hurt due to over-smoothing; ACM is reported as a slight exception. Parameter analysis further indicates that optimal fusion weights vary by dataset: ACM favors a larger GCN weight, Citeseer a larger AE weight, HHAR and Cora a larger Transformer weight, and Reuters a strong GCN weight plus substantial Transformer contribution.

6. Related graph-fusion research, limitations, and applications

Tri-GFN belongs to a broader family of graph-fusion methods, but its design target is distinct. SF-GCN learns a fused adjacency

$D_{ii}=\sum_j (A+I)_{ij},$9

for multi-view semi-supervised classification by balancing specificity and commonality in spectral space (Lin et al., 2019). LGCN-FF jointly learns feature fusion and graph fusion, then refines the fused adjacency with the Differentiable Shrinkage Activation

$E=[s;t]\in\mathbb{R}^{2\times \delta},$0

rather than relying only on a weighted graph sum (Chen et al., 2022). IGNet, in a different application domain, uses a dual-branch graph interaction module with cross-modality message passing and leader nodes for infrared-visible image fusion (Li et al., 2023). Relative to these models, Tri-GFN is distinguished by its explicit three-branch coupling among AE, GCN, and Graph Transformer for unsupervised attributed graph clustering (Li et al., 18 Jul 2025).

Several limitations are explicit. The method has substantial hyperparameter sensitivity, including $E=[s;t]\in\mathbb{R}^{2\times \delta},$1, $E=[s;t]\in\mathbb{R}^{2\times \delta},$2, $E=[s;t]\in\mathbb{R}^{2\times \delta},$3, and $E=[s;t]\in\mathbb{R}^{2\times \delta},$4. Training efficiency remains a concern on large-scale graphs, and adding more propagation layers can degrade performance through over-smoothing. A further issue is textual ambiguity in the manuscript itself: some equations, especially those related to reconstruction and certain coefficient symbols, are imprecisely typeset. A plausible implication is that faithful reimplementation depends on careful reconciliation of equations, figures, and parameter tables rather than on any single equation block.

The paper connects Tri-GFN’s performance on Reuters to automatic news classification and topic retrieval (Li et al., 18 Jul 2025). It also situates the model more broadly in application areas such as social network analysis, bioinformatics, recommendation systems, and network security. Those use cases follow directly from the model’s intended operating regime: graphs with rich node attributes, meaningful local structure, and nontrivial long-range dependencies.