AGREE: Any-Type Graph Representation Learning

• AGREE is an end-to-end framework for unsupervised clustering on heterogeneous attributed graphs, integrating arbitrary information sources into a unified embedding space.
• It employs multi-level attribute alignment and quaternion graph convolution to mitigate over-smoothing and over-dominating issues while avoiding predefined cluster counts.
• Empirical evaluations on 19 datasets show that AGREE outperforms 13 baselines in clustering accuracy and robustness across both attributed graph and mixed-type settings.

Any-type attributed Graph REpresentation lEarning (AGREE) is an end-to-end framework for unsupervised clustering on attributed graphs with heterogeneous attributes, and, in a broader methodological sense, a graph representation-learning perspective in which embeddings are organized around attribute-grounded types or arbitrary information sources rather than fixed node identities. In its 2026 formulation, AGREE accepts either an attributed graph $G=\{\mathbf{A},\mathbf{X}\}$ or a mixed-type attribute dataset $$\mathcal{H}=\langle \mathcal{X},\mathcal{A},\mathcal{O}\rangle$$, unifies attributed graphs and any-type attributed data through multi-level alignment and similarity-based graph construction, applies shallow quaternion graph convolution, and jointly optimizes graph reconstruction and clustering without requiring a predefined number of clusters during training (Chen et al., 22 Jun 2026). Earlier work anticipated this “any-type” viewpoint by replacing node-identity walks with attributed random walks over types (Ahmed et al., 2017) and by integrating arbitrary information sources as auxiliary weighted graphs linked through adaptive transition relations (Qin, 2023).

1. Conceptual lineage and the meaning of “any-type”

The modern AGREE formulation sits at the intersection of two earlier lines of research. The first is inductive representation learning on attributed graphs, where the central move is to replace node-ID-centric random walks with a mapping $\phi:\mathbf{x}\rightarrow w$ from node attributes to discrete types, and then learn embeddings over type sequences rather than node identities. In that setting, an attributed walk of length $\ell$ is defined as a sequence of adjacent node types $$\phi(\mathbf{x}_{i_1}),\phi(\mathbf{x}_{i_2}),\ldots,\phi(\mathbf{x}_{i_{\ell+1}})$$ associated with adjacent nodes in $G$, so multiple nodes can share a type and thus share an embedding (Ahmed et al., 2017). This directly addresses three limitations identified for identity-based random-walk methods such as DeepWalk and node2vec: they are inherently transductive, not space efficient, and generally lack support for attributed graphs. The same work reports that the type-based formulation requires on average $853\times$ less space and is “accurate with an average improvement of $16.1\%$ across a variety of graphs from several domains” on link prediction (Ahmed et al., 2017).

The second line is arbitrary-source integration in attributed graphs. AHGR models each available information source $I_l$ as an auxiliary weighted graph $G_l=(V,E_l)$ with adjacency $$\mathcal{H}=\langle \mathcal{X},\mathcal{A},\mathcal{O}\rangle$$0, so high-order proximities, community structure, node attributes, and other sources can be represented in a common reweighting formalism. A shared latent representation $$\mathcal{H}=\langle \mathcal{X},\mathcal{A},\mathcal{O}\rangle$$1 is then linked to source-specific basic embeddings $$\mathcal{H}=\langle \mathcal{X},\mathcal{A},\mathcal{O}\rangle$$2 through transition matrices $$\mathcal{H}=\langle \mathcal{X},\mathcal{A},\mathcal{O}\rangle$$3, with the objective $$\mathcal{H}=\langle \mathcal{X},\mathcal{A},\mathcal{O}\rangle$$4 and a consistency indicator $$\mathcal{H}=\langle \mathcal{X},\mathcal{A},\mathcal{O}\rangle$$5 quantifying how explicitly each source aligns with the shared space (Qin, 2023).

Taken together, these precursors make the phrase “any-type” precise in two complementary ways. First, “type” may mean attribute-derived node categories, structural roles, communities, or hybrid semantics. Second, “type” may also refer to arbitrary information channels that can be cast as graph-like relations and then integrated adaptively. The 2026 AGREE framework specializes this broader view to heterogeneous attributed graph clustering through aligned feature construction, similarity-based graph induction, and quaternion graph representation learning (Chen et al., 22 Jun 2026).

2. Formal problem setting and the central challenges

AGREE addresses unsupervised clustering on attributed graphs whose nodes possess structural relations and mixed-type attributes. For an attributed graph, the input is

$$\mathcal{H}=\langle \mathcal{X},\mathcal{A},\mathcal{O}\rangle$$6

where $$\mathcal{H}=\langle \mathcal{X},\mathcal{A},\mathcal{O}\rangle$$7 is the adjacency matrix and $$\mathcal{H}=\langle \mathcal{X},\mathcal{A},\mathcal{O}\rangle$$8 is the node attribute matrix, with each row a $$\mathcal{H}=\langle \mathcal{X},\mathcal{A},\mathcal{O}\rangle$$9-dimensional attribute vector. For more general mixed-type attribute data not initially given as a graph, the framework defines

$\phi:\mathbf{x}\rightarrow w$0

where $\phi:\mathbf{x}\rightarrow w$1 is the object set, $\phi:\mathbf{x}\rightarrow w$2 is the feature set, and $\phi:\mathbf{x}\rightarrow w$3 specifies the possible values for each categorical feature. Features are partitioned into categorical and numerical subsets, $\phi:\mathbf{x}\rightarrow w$4 and $\phi:\mathbf{x}\rightarrow w$5, with $\phi:\mathbf{x}\rightarrow w$6 (Chen et al., 22 Jun 2026).

The clustering objective is to partition nodes into $\phi:\mathbf{x}\rightarrow w$7 groups by jointly exploiting topology and attributes, or equivalently to learn node embeddings that are clustering-friendly. The paper identifies three technical obstacles. The first is heterogeneous attribute unification: numerical features admit Euclidean-style distances, whereas categorical features require representations that reflect co-occurrence statistics and inter-feature dependencies rather than mere one-hot orthogonality. The second is graph construction: when the topology is noisy, incomplete, or absent, the induced graph from attributes must preserve cluster structure rather than simply reflect high-dimensional scale effects. The third is representation degradation during graph learning, described through two phenomena: over-smoothing (OS) and over-dominating (OD) (Chen et al., 22 Jun 2026).

OS refers to the tendency of repeated graph propagation to homogenize node representations. AGREE uses the normalized adjacency

$\phi:\mathbf{x}\rightarrow w$8

and notes that repeated applications of real-valued graph propagation drive embeddings toward excessive similarity, particularly on dense similarity graphs. OD is introduced as a distinct effect in graph clustering: topological influence can dominate attribute information, so nodes with similar connectivity but different attributes become overly similar in the learned space. A recurring misconception is that AGREE is merely a graph-clustering model for already-given homogeneous graphs; in fact, the framework explicitly covers mixed-type tabular data without an initial graph and treats resistance to OD as a primary design objective (Chen et al., 22 Jun 2026).

3. Multi-level alignment and similarity-based graph construction

AGREE’s “any-type” mechanism begins with aligned encoding of heterogeneous attributes. For a categorical feature $\phi:\mathbf{x}\rightarrow w$9 with value set $\ell$0, the framework first defines value-level occurrence probabilities

$\ell$1

capturing global frequency information. It then introduces feature-level encoding through conditional probability distributions

$\ell$2

which encode how a value in feature $\ell$3 co-occurs with values of other categorical features. To bridge categorical and numerical attribute types, AGREE further computes distances between categorical values,

$\ell$4

and uses many one-dimensional projection spaces $\ell$5 to encode within-feature geometry. These components are concatenated into the final categorical encoding

$\ell$6

All encoded categorical features are combined with the numerical features to form the aligned feature set $\ell$7 and the aligned object representation $\ell$8 (Chen et al., 22 Jun 2026).

Object-level alignment is then defined through a unified mixed-type dissimilarity. For two objects $\ell$9,

$$\phi(\mathbf{x}_{i_1}),\phi(\mathbf{x}_{i_2}),\ldots,\phi(\mathbf{x}_{i_{\ell+1}})$$0

where

$$\phi(\mathbf{x}_{i_1}),\phi(\mathbf{x}_{i_2}),\ldots,\phi(\mathbf{x}_{i_{\ell+1}})$$1

Here the categorical dissimilarity is

$$\phi(\mathbf{x}_{i_1}),\phi(\mathbf{x}_{i_2}),\ldots,\phi(\mathbf{x}_{i_{\ell+1}})$$2

Numerical features are discretized into 5 equal-width intervals for the dependency computations, but the final per-feature numerical distance remains $$\phi(\mathbf{x}_{i_1}),\phi(\mathbf{x}_{i_2}),\ldots,\phi(\mathbf{x}_{i_{\ell+1}})$$3. This distinction is important because it preserves numerical magnitude while allowing a common statistical framework for cross-type alignment (Chen et al., 22 Jun 2026).

The graph is then constructed as a fully connected graph with

$$\phi(\mathbf{x}_{i_1}),\phi(\mathbf{x}_{i_2}),\ldots,\phi(\mathbf{x}_{i_{\ell+1}})$$4

followed by normalization through $$\phi(\mathbf{x}_{i_1}),\phi(\mathbf{x}_{i_2}),\ldots,\phi(\mathbf{x}_{i_{\ell+1}})$$5. The paper explicitly remarks that using this graph-based unified dissimilarity, rather than Euclidean distance on the expanded aligned attributes, prevents expanded categorical encodings from dominating the geometry merely because of dimensionality. A common point of confusion is that $$\phi(\mathbf{x}_{i_1}),\phi(\mathbf{x}_{i_2}),\ldots,\phi(\mathbf{x}_{i_{\ell+1}})$$6 encodes dissimilarity rather than a conventional similarity kernel; AGREE nonetheless treats this matrix directly as adjacency and normalizes it before graph propagation (Chen et al., 22 Jun 2026).

4. Quaternion graph representation learning

After alignment and graph construction, AGREE applies Four-View Projection (FVP) to map $$\phi(\mathbf{x}_{i_1}),\phi(\mathbf{x}_{i_2}),\ldots,\phi(\mathbf{x}_{i_{\ell+1}})$$7 into four learned views: $$\phi(\mathbf{x}_{i_1}),\phi(\mathbf{x}_{i_2}),\ldots,\phi(\mathbf{x}_{i_{\ell+1}})$$8 These are assembled into a quaternion feature matrix

$$\phi(\mathbf{x}_{i_1}),\phi(\mathbf{x}_{i_2}),\ldots,\phi(\mathbf{x}_{i_{\ell+1}})$$9

where quaternion multiplication is implemented by the Hamilton product. The paper notes that the quaternion transformation has $G$0 degrees of freedom compared with a real-valued equivalent at the same parameter scale, because the four real-valued parameter blocks induce 16 structured transformation blocks in the real implementation (Chen et al., 22 Jun 2026).

Graph propagation is performed by Quaternion Graph Encoders (QGE). With $G$1, each layer is

$G$2

where $G$3 denotes the Hamilton product, $G$4 is the normalized adjacency, and $G$5 is an activation function such as ReLU applied component-wise. The final quaternion representation $G$6 is fused into a real embedding

$G$7

where $G$8 is implemented as averaging over the real and three imaginary components. Graph reconstruction uses

$G$9

The architecture is deliberately shallow, typically two QGE layers, with the expressiveness of quaternion coupling intended to reduce the need for deeper propagation and thereby mitigate OS while strengthening attribute-side modeling against OD (Chen et al., 22 Jun 2026).

The training objective is

$853\times$0

where $853\times$1 is a KL-style reconstruction loss between $853\times$2 and $853\times$3,

$853\times$4

and the clustering term is the relaxed spectral objective

$853\times$5

Because this loss does not introduce explicit cluster centers or assignment distributions, AGREE does not require the number of clusters $853\times$6 during training. After training, one computes the $853\times$7 smallest-eigenvalue eigenvectors of $853\times$8 and runs $853\times$9-means on them to obtain final cluster labels (Chen et al., 22 Jun 2026).

5. Empirical behavior and comparative position

AGREE is evaluated on 19 datasets: 9 attributed graph datasets and 10 mixed-type attributed datasets. The attributed graph collection includes ACM, Wiki-CS, CiteSeer, DBLP, FILM, Cora, Wisconsin, USA Air-Traffic, and Amazon Photo. The mixed-type collection includes Mammographic, Heart Failure, Breast Cancer, Autism-Adolescent, Tic-Tac-Toe, Zoo, Yeast, Glass, Wine, and Iris. The baselines span classical clustering, graph autoencoders, contrastive graph clustering, and explicit structure-learning methods, for a total of 13 methods including k-means, GAE, ARVGAE, DAEGC, DFCN, EGAE, CCGC, CONVERT, SCDGN, MAGI, GLAC, CDC, and DESE. Evaluation uses external metrics—Accuracy (ACC), Normalized Mutual Information (NMI), and Adjusted Rand Index (ARI)—and internal metrics—Silhouette Coefficient (SC), Davies–Bouldin Index (DBI), and Calinski–Harabasz Index (CHI). Each experiment is repeated 10 times and reported as mean $16.1\%$0 standard deviation (Chen et al., 22 Jun 2026).

On mixed-type attributed datasets, AGREE achieves the best average rank, approximately $16.1\%$1, substantially ahead of the listed baselines. Representative results include Mammographic ARI $16.1\%$2 for AGREE versus $16.1\%$3 for DESE and $16.1\%$4 for k-means; Tic-Tac-Toe ARI $16.1\%$5 for AGREE versus $16.1\%$6 for CDC and approximately $16.1\%$7–$16.1\%$8 for several other baselines; Wine ACC $16.1\%$9 for AGREE versus $I_l$0 for CDC and $I_l$1 for GLAC; and Iris ACC $I_l$2 for AGREE versus $I_l$3 for SCDGN and $I_l$4 for k-means. On attributed graph datasets, AGREE attains the best average rank, approximately $I_l$5, with examples such as ACM ACC/NMI/ARI $I_l$6 and Cora ACC/NMI/ARI $I_l$7 (Chen et al., 22 Jun 2026).

Ablation studies isolate three components: alignment, quaternion representation learning, and the spectral objective. On attributed graph datasets, the full model outperforms variants such as a real-valued MLP plus 2-layer GCN baseline, a variant without FVP, a complex-valued alternative to QGE, and a version without the spectral loss, dominating in 24–26 of 27 reported metrics. On mixed-type datasets, alignment alone does not consistently improve performance, quaternion representation learning alone improves several datasets, and the full combination yields the best or near-best results on almost all metrics. Additional depth studies show that both real-valued and quaternion encoders degrade as layer count increases, but AGREE-QGE degrades more slowly; topology-perturbation experiments based on injecting cross-class edges show higher CKA similarity and slower ARI degradation for AGREE-QGE than for a real-valued counterpart, which the paper interprets as evidence of reduced OD and better robustness to topology noise (Chen et al., 22 Jun 2026).

In relation to prior graph representation-learning paradigms, AGREE differs from structure-only methods such as DeepWalk, node2vec, LINE, SDNE, GraRep, and AROPE by explicitly modeling attributes; it differs from classical structure-plus-attribute factorization methods such as TADW, AANE, and FSCNMF by directly targeting heterogeneous feature alignment and by treating the graph itself as constructible from mixed-type data; and it differs from arbitrary-source integration methods such as AHGR by instantiating a single end-to-end quaternion autoencoder rather than a two-stage source-wise factorization architecture (Qin, 2023).

6. Limitations, misconceptions, and future directions

AGREE’s main limitations are computational and structural. The alignment stage includes a worst-case $I_l$8 term for attribute-type encoding and an $I_l$9 term for object-level distances, where $G_l=(V,E_l)$0. The overall time complexity is reported as

$G_l=(V,E_l)$1

and memory is dominated by the dense $G_l=(V,E_l)$2 adjacency. The framework is therefore characterized as more suitable for small-to-medium graphs unless approximations are introduced (Chen et al., 22 Jun 2026).

Several common misconceptions are explicitly ruled out by the formulation. AGREE is not a method that learns cluster centers during training; cluster assignments are not updated during optimization, and clustering is performed only after training via classical spectral clustering on the reconstructed graph. It is not “$G_l=(V,E_l)$3-free” in the sense of eliminating the need to choose a number of clusters at deployment time; rather, it is independent of $G_l=(V,E_l)$4 during representation learning and can be reused across different $G_l=(V,E_l)$5 values without retraining. It is also not yet a native solution for heterogeneous nodes, heterogeneous edges, or dynamic graphs; these are identified as future extensions rather than current capabilities (Chen et al., 22 Jun 2026).

The future directions named for AGREE include extending the framework to more complex graph forms such as heterogeneous nodes or edges and dynamic graphs, applying it in federated graph learning settings, and studying debiasing under skewed or imbalanced attributes (Chen et al., 22 Jun 2026). Earlier AGREE-like work suggests additional trajectories: end-to-end differentiable mappings from attributes to types, hierarchical or multi-view type spaces, and tighter integration with GNN-based encoders or arbitrary-source fusion schemes (Ahmed et al., 2017). A plausible implication is that AGREE can be understood less as a single architecture than as a research program centered on type construction, source alignment, and inductive or clustering-oriented graph representation learning across heterogeneous attribute regimes.