Attentive Graph Clustering Network (AGCN)

• Attentive Graph Clustering Network (AGCN) is an unsupervised framework that leverages attention to combine node attributes and graph topology for enhanced clustering.
• It employs heterogeneity-wise and scale-wise fusion modules to dynamically balance feature sources and propagation scales, addressing multi-scale graph information.
• Empirical evaluations demonstrate that AGCN consistently outperforms baseline models in metrics such as ACC, NMI, ARI, and F1 across diverse benchmark datasets.

Attentive Graph Clustering Network (AGCN) denotes a family of unsupervised graph clustering models in which attention is used to regulate how node attributes, topological signals, propagation scales, or cluster-level distributions contribute to representation learning and cluster assignment. In its most direct 2021 formulation, AGCN is a deep clustering framework that jointly learns node representations and cluster assignments from both node attributes and graph structure through a heterogeneity-wise fusion module and a scale-wise fusion module, under a unified unsupervised optimization objective (Peng et al., 2021). Closely related formulations include attentional graph autoencoding with self-training (Wang et al., 2019), adaptive high-order graph convolution with graph-specific scale selection (Zhang et al., 2019), dual self-supervised attention-guided fusion models that explicitly extend AGCN (Peng et al., 2021), and structure-aware Transformer variants that reinterpret graph clustering as masked attention over high-order neighborhoods (Xie et al., 18 Sep 2025).

1. Problem setting and conceptual scope

AGCN is defined on graph-structured data with node attributes. In the canonical attributed-graph setting, the data are represented by a node attribute matrix $\mathbf{X} \in \mathbb{R}^{n \times d}$ and an adjacency matrix $\mathbf{A} \in \mathbb{R}^{n \times n}$, with the clustering task being: given $(\mathbf{A}, \mathbf{X})$ and the number of clusters $k$, learn an embedding and assign each node to a cluster without labels (Peng et al., 2021). The central motivation is that deep clustering methods based on an auto-encoder and a graph convolutional network often fuse attribute and structural information inflexibly and frequently use only the deepest layer representation, thereby overlooking the multi-scale information embedded at different layers (Peng et al., 2021).

Within the broader literature, AGCN is best understood as a graph clustering architecture in which attention modulates one or more of four objects: heterogeneous feature sources, neighborhood scales, neighbors or propagation depths, and cluster-level probability distributions. DAEGC, for example, is a goal-directed deep learning approach for attributed graphs in which an attention network captures the importance of neighboring nodes to a target node and a self-training graph clustering process iteratively refines cluster assignments (Wang et al., 2019). AGC, while not explicitly called AGCN, provides an adaptive graph convolution view in which selecting the appropriate order of graph filtering functions as a coarse-grained attention over propagation depth or scale (Zhang et al., 2019).

2. Canonical architecture of the 2021 AGCN

The 2021 “Attention-driven Graph Clustering Network” combines an auto-encoder, a GCN stack, and two attention-driven fusion modules: AGCN-H for heterogeneity-wise fusion and AGCN-S for scale-wise fusion (Peng et al., 2021). The auto-encoder learns node attribute representations by minimizing a reconstruction loss,

$$\mathcal{L}_{R} = \left\| \mathbf{X} - \hat{\mathbf{X}} \right\|^2_F,$$

with encoder and decoder layers

$$\mathbf{H}_{i} = \phi \left( \mathbf{W}_{i}^{e}\mathbf{H}_{i-1} + \mathbf{b}_{i}^{e} \right), \qquad \hat{\mathbf{H}}_{i} = \phi \left( \mathbf{W}_{i}^{d}\hat{\mathbf{H}}_{i-1} + \mathbf{b}_{i}^{d} \right).$$

The final latent AE feature is $\mathbf{H}_l$ (Peng et al., 2021).

The GCN pathway starts from $\mathbf{Z}_0 = \mathbf{X}$ and propagates fused features through the normalized adjacency:

$$\mathbf{Z}_{i+1} = \mathrm{LeakyReLU}\left( \mathbf{D}^{-\frac{1}{2}} (\mathbf{A} + \mathbf{I}) \mathbf{D}^{-\frac{1}{2}} \mathbf{Z}_i' \mathbf{W}_i \right).$$

Here $\mathbf{Z}_i'$ is not the raw GCN feature but the output of heterogeneity-wise fusion between the GCN feature $\mathbf{A} \in \mathbb{R}^{n \times n}$0 and the AE feature $\mathbf{A} \in \mathbb{R}^{n \times n}$1 at the same depth (Peng et al., 2021).

The heterogeneity-wise fusion module dynamically balances attribute and topology. For each layer $\mathbf{A} \in \mathbb{R}^{n \times n}$2, AGCN concatenates $\mathbf{A} \in \mathbb{R}^{n \times n}$3 and computes a two-dimensional attention vector:

$\mathbf{A} \in \mathbb{R}^{n \times n}$4

with $\mathbf{A} \in \mathbb{R}^{n \times n}$5. The fused feature is then

$\mathbf{A} \in \mathbb{R}^{n \times n}$6

This makes the contribution of the node attribute feature and the topological graph feature node-specific and layer-specific rather than fixed (Peng et al., 2021).

The scale-wise fusion module addresses the fact that clustering performance may depend on intermediate as well as deepest layers. AGCN concatenates the multi-scale features $\mathbf{A} \in \mathbb{R}^{n \times n}$7 with $\mathbf{A} \in \mathbb{R}^{n \times n}$8, computes attention across scales,

$\mathbf{A} \in \mathbb{R}^{n \times n}$9

and forms a weighted multi-scale representation

$(\mathbf{A}, \mathbf{X})$0

The result is a clustering representation that preserves multiple receptive-field scales while down-weighting noisy or over-smoothed layers (Peng et al., 2021).

3. Clustering objective and optimization

The canonical AGCN couples representation learning and clustering by using a DEC-style self-training objective together with reconstruction. Cluster centers $(\mathbf{A}, \mathbf{X})$1 are initialized by K-means on the final AE embedding $(\mathbf{A}, \mathbf{X})$2, and soft assignments are computed with Student’s $(\mathbf{A}, \mathbf{X})$3-distribution:

$(\mathbf{A}, \mathbf{X})$4

An auxiliary target distribution sharpens these assignments,

$(\mathbf{A}, \mathbf{X})$5

and the clustering loss aligns both the graph-based prediction $(\mathbf{A}, \mathbf{X})$6 and the AE-based distribution $(\mathbf{A}, \mathbf{X})$7 to this target:

$(\mathbf{A}, \mathbf{X})$8

The total objective is

$(\mathbf{A}, \mathbf{X})$9

A final graph-aware prediction layer produces soft cluster assignments,

$k$0

and hard labels are given by $k$1 (Peng et al., 2021).

The training protocol is two-stage. The AE is pre-trained for 30 epochs on $k$2 with learning rate $k$3, after which the whole model is jointly optimized. In the reported experiments, both AE and GCN hidden dimensions are set to $k$4-$k$5-$k$6-$k$7, the number of layers is $k$8, batch size is $k$9, and training is repeated 10 times with mean $$\mathcal{L}_{R} = \left\| \mathbf{X} - \hat{\mathbf{X}} \right\|^2_F,$$0 standard deviation reported (Peng et al., 2021).

This unified optimization principle reappears in related attentive graph clustering models. DAEGC uses an attentional encoder, an inner product decoder, Student’s $$\mathcal{L}_{R} = \left\| \mathbf{X} - \hat{\mathbf{X}} \right\|^2_F,$$1-distribution for soft assignments, a target distribution $$\mathcal{L}_{R} = \left\| \mathbf{X} - \hat{\mathbf{X}} \right\|^2_F,$$2, and a joint objective $$\mathcal{L}_{R} = \left\| \mathbf{X} - \hat{\mathbf{X}} \right\|^2_F,$$3, thereby making the embedding explicitly goal-directed for clustering rather than only reconstruction-driven (Wang et al., 2019).

4. Related attentive graph clustering paradigms

The AGCN lineage contains several distinct but technically adjacent formulations. DAEGC is a representative attentional embedding approach for attributed graph clustering: it replaces uniform aggregation with topology-aware neighbor attention, reconstructs adjacency through an inner-product decoder, and iteratively refines cluster assignments through KL divergence between current and sharpened assignment distributions (Wang et al., 2019). Its encoder uses a proximity matrix

$$\mathcal{L}_{R} = \left\| \mathbf{X} - \hat{\mathbf{X}} \right\|^2_F,$$4

so attention depends jointly on feature content and higher-order topology.

AGC offers a different route to attentiveness. It uses a nonparametric high-order graph filter

$$\mathcal{L}_{R} = \left\| \mathbf{X} - \hat{\mathbf{X}} \right\|^2_F,$$5

with frequency response

$$\mathcal{L}_{R} = \left\| \mathbf{X} - \hat{\mathbf{X}} \right\|^2_F,$$6

and selects the order $$\mathcal{L}_{R} = \left\| \mathbf{X} - \hat{\mathbf{X}} \right\|^2_F,$$7 adaptively by stopping at the first local minimum of the intra-cluster distance. Although AGC has no explicit attention weights, its adaptivity is mathematically close to attention over propagation depth or graph scale (Zhang et al., 2019).

DAGC is explicitly presented as the journal extension of the AGCN conference version. It preserves heterogeneity-wise fusion and scale-wise fusion, adds distribution-wise fusion to combine AE-based and graph-based clustering distributions, and introduces dual self-supervision: a soft self-supervision strategy with a triplet KL divergence loss and a hard self-supervision strategy with a pseudo supervision loss (Peng et al., 2021). In this formulation, attention operates at three levels: AE-versus-GCN fusion, multi-scale fusion across layers, and fusion of clustering distributions.

Subsequent work broadens the architectural space further. GraphHAM formulates clustering as hierarchical attentive membership, learning node embeddings and latent group embeddings jointly, with both group-level and individual-level attentions and explicit structural constraints over inferred memberships (Lin et al., 2021). RCLG frames attentive graph clustering as adaptive local-global integration, where local multi-depth propagation features are fused by attention and global semantic prototypes derived from evolving cluster centers are aggregated through attention under a dual-view contrastive objective (Zhang et al., 27 May 2026). CLATT is not an unsupervised clustering model, but it introduces cluster attention as a reusable module: nodes are divided into clusters with off-the-shelf graph community detection algorithms, and each node attends to all other nodes in each cluster, which suggests a direct path toward hierarchical or end-to-end cluster-aware AGCNs (Platonov et al., 8 Apr 2026).

A distinct 2025 model again adopts the name AGCN but changes the backbone completely. It is a structure-aware Transformer for unsupervised node clustering in which attention is explicitly masked by $$\mathcal{L}_{R} = \left\| \mathbf{X} - \hat{\mathbf{X}} \right\|^2_F,$$8-hop neighborhoods from $$\mathcal{L}_{R} = \left\| \mathbf{X} - \hat{\mathbf{X}} \right\|^2_F,$$9, combines a KV cache mechanism with a pairwise margin contrastive loss, and uses k-means on the final embeddings for clustering (Xie et al., 18 Sep 2025). In that model, AGCN no longer denotes AE-plus-GCN fusion; it denotes a fully attention-driven clustering architecture that constrains self-attention by graph structure.

5. Empirical behavior and reported performance

The 2021 AGCN is evaluated on six benchmark datasets: USPS, HHAR, Reuters, ACM, CiteSeer, and DBLP, using ACC, NMI, ARI, and macro F1 as metrics (Peng et al., 2021). It consistently outperforms AE, DEC, IDEC, GAE, VGAE, DAEGC, ARGA, and SDCN. On HHAR, for example, SDCN reports ACC $$\mathbf{H}_{i} = \phi \left( \mathbf{W}_{i}^{e}\mathbf{H}_{i-1} + \mathbf{b}_{i}^{e} \right), \qquad \hat{\mathbf{H}}_{i} = \phi \left( \mathbf{W}_{i}^{d}\hat{\mathbf{H}}_{i-1} + \mathbf{b}_{i}^{d} \right).$$0, NMI $$\mathbf{H}_{i} = \phi \left( \mathbf{W}_{i}^{e}\mathbf{H}_{i-1} + \mathbf{b}_{i}^{e} \right), \qquad \hat{\mathbf{H}}_{i} = \phi \left( \mathbf{W}_{i}^{d}\hat{\mathbf{H}}_{i-1} + \mathbf{b}_{i}^{d} \right).$$1, ARI $$\mathbf{H}_{i} = \phi \left( \mathbf{W}_{i}^{e}\mathbf{H}_{i-1} + \mathbf{b}_{i}^{e} \right), \qquad \hat{\mathbf{H}}_{i} = \phi \left( \mathbf{W}_{i}^{d}\hat{\mathbf{H}}_{i-1} + \mathbf{b}_{i}^{d} \right).$$2, and F1 $$\mathbf{H}_{i} = \phi \left( \mathbf{W}_{i}^{e}\mathbf{H}_{i-1} + \mathbf{b}_{i}^{e} \right), \qquad \hat{\mathbf{H}}_{i} = \phi \left( \mathbf{W}_{i}^{d}\hat{\mathbf{H}}_{i-1} + \mathbf{b}_{i}^{d} \right).$$3, whereas AGCN reports ACC $$\mathbf{H}_{i} = \phi \left( \mathbf{W}_{i}^{e}\mathbf{H}_{i-1} + \mathbf{b}_{i}^{e} \right), \qquad \hat{\mathbf{H}}_{i} = \phi \left( \mathbf{W}_{i}^{d}\hat{\mathbf{H}}_{i-1} + \mathbf{b}_{i}^{d} \right).$$4, NMI $$\mathbf{H}_{i} = \phi \left( \mathbf{W}_{i}^{e}\mathbf{H}_{i-1} + \mathbf{b}_{i}^{e} \right), \qquad \hat{\mathbf{H}}_{i} = \phi \left( \mathbf{W}_{i}^{d}\hat{\mathbf{H}}_{i-1} + \mathbf{b}_{i}^{d} \right).$$5, ARI $$\mathbf{H}_{i} = \phi \left( \mathbf{W}_{i}^{e}\mathbf{H}_{i-1} + \mathbf{b}_{i}^{e} \right), \qquad \hat{\mathbf{H}}_{i} = \phi \left( \mathbf{W}_{i}^{d}\hat{\mathbf{H}}_{i-1} + \mathbf{b}_{i}^{d} \right).$$6, and F1 $$\mathbf{H}_{i} = \phi \left( \mathbf{W}_{i}^{e}\mathbf{H}_{i-1} + \mathbf{b}_{i}^{e} \right), \qquad \hat{\mathbf{H}}_{i} = \phi \left( \mathbf{W}_{i}^{d}\hat{\mathbf{H}}_{i-1} + \mathbf{b}_{i}^{d} \right).$$7 (Peng et al., 2021). On DBLP, SDCN reports ACC $$\mathbf{H}_{i} = \phi \left( \mathbf{W}_{i}^{e}\mathbf{H}_{i-1} + \mathbf{b}_{i}^{e} \right), \qquad \hat{\mathbf{H}}_{i} = \phi \left( \mathbf{W}_{i}^{d}\hat{\mathbf{H}}_{i-1} + \mathbf{b}_{i}^{d} \right).$$8 and F1 $$\mathbf{H}_{i} = \phi \left( \mathbf{W}_{i}^{e}\mathbf{H}_{i-1} + \mathbf{b}_{i}^{e} \right), \qquad \hat{\mathbf{H}}_{i} = \phi \left( \mathbf{W}_{i}^{d}\hat{\mathbf{H}}_{i-1} + \mathbf{b}_{i}^{d} \right).$$9, whereas AGCN reports ACC $\mathbf{H}_l$0 and F1 $\mathbf{H}_l$1 (Peng et al., 2021). The ablation results further show that AGCN-H, naive scale fusion, and attention-based scale fusion each contribute, with the full model performing best and scale attention helping suppress over-smoothed intermediate layers on HHAR (Peng et al., 2021).

Earlier attentional clustering models also report strong gains. On Cora, DAEGC reports ACC $\mathbf{H}_l$2, NMI $\mathbf{H}_l$3, F-score $\mathbf{H}_l$4, and ARI $\mathbf{H}_l$5, outperforming VGAE and other baselines listed in the study; similar improvements are reported on Citeseer and Pubmed (Wang et al., 2019). DAGC later reports that it consistently outperforms state-of-the-art methods on nine datasets and improves the ARI by more than $\mathbf{H}_l$6 over the best baseline, with the largest cited gain occurring on DBLP relative to SDCN (Peng et al., 2021).

The 2025 structure-aware Transformer AGCN reports state-of-the-art ACC and NMI on all seven heterophilic datasets in its evaluation. On Cornell it reports ACC $\mathbf{H}_l$7 and NMI $\mathbf{H}_l$8, compared with the cited best baseline DGCN at ACC $\mathbf{H}_l$9 and NMI $\mathbf{Z}_0 = \mathbf{X}$0; on Roman-Empire it reports ACC $\mathbf{Z}_0 = \mathbf{X}$1 and NMI $\mathbf{Z}_0 = \mathbf{X}$2 (Xie et al., 18 Sep 2025). On homophilic datasets, it reports the best performance on Cora, Citeseer, and UAT, and second-best on Pubmed, while also outperforming a variant in which its encoder is replaced by a GCN on large graphs such as Flickr and Twitch-Gamers (Xie et al., 18 Sep 2025). These results suggest that attention in graph clustering is not intrinsically redundant; rather, its utility depends strongly on how graph structure constrains the attention space.

6. Terminology, misconceptions, and scope boundaries

A common source of confusion is that “AGCN” is not a unique acronym across graph learning. In graph clustering, the term usually refers to architectures that jointly learn graph embeddings and cluster assignments with attention over nodes, edges, scales, or feature sources (Peng et al., 2021). Outside clustering, however, the same acronym has been used for different objectives. CP-AGCN is a “Pytorch-based attention-informed graph convolutional network” for binary classification of normal versus abnormal infant movements indicating risk of cerebral palsy; it learns graph representations of infant skeletal motion in time and frequency and uses attention to weight joints or frequency bins, but it is explicitly not a clustering model (Zhang et al., 2022). Likewise, “Attention-based Graph Convolution Networks” for point clouds is a supervised architecture for 3D classification and segmentation, using attention-based neighbor aggregation on KNN graphs rather than an unsupervised clustering objective (Xie et al., 2019).

Another boundary case is “attention graph clustering” inside larger systems. GCAGC for co-saliency detection includes an attention graph clustering module that optimizes a weighted kernel $\mathbf{Z}_0 = \mathbf{X}$3-means–like objective,

$\mathbf{Z}_0 = \mathbf{X}$4

with $\mathbf{Z}_0 = \mathbf{X}$5, and uses the resulting continuous foreground indicator as a co-attention map in an encoder–graph–clustering–decoder pipeline (Zhang et al., 2020). This is graph clustering with attention, but in service of co-saliency detection rather than generic node clustering.

The literature therefore supports a narrow and a broad usage. In the narrow usage, AGCN denotes the 2021 attention-driven deep clustering network and its immediate extension DAGC (Peng et al., 2021, Peng et al., 2021). In the broad usage, AGCN denotes any clustering-oriented graph architecture in which attention governs heterogeneous fusion, neighborhood relevance, scale selection, prototype interaction, or cluster-restricted communication (Wang et al., 2019, Xie et al., 18 Sep 2025, Platonov et al., 8 Apr 2026). A plausible implication is that the term should always be interpreted relative to the training objective and the specific attention locus, rather than from the acronym alone.