uAGNN: Unsupervised Antisymmetric GNN

• The paper introduces uAGNN, a novel unsupervised community detection method that uses antisymmetric propagation to preserve long-range information in heterophilic graphs.
• It models graph propagation as a non-dissipative dynamical system via ODE formulation and forward Euler discretization, ensuring stable information flow.
• Empirical results demonstrate that uAGNN outperforms traditional methods on low- and medium-homophily datasets by balancing local aggregation with global reconstruction.

Searching arXiv for the exact uAGNN paper and closely related AGNN work to ground the article in current arXiv records. The Unsupervised Antisymmetric Graph Neural Network (uAGNN) is an unsupervised graph clustering method for community detection that is designed for graphs in which informative relationships may be non-local rather than concentrated in immediate neighborhoods. In the formulation reported for uAGNN, graph propagation is cast as a non-dissipative dynamical system, with antisymmetric weight matrices used to preserve stability while propagating long-range information effectively (Leeney et al., 15 Aug 2025). The method is situated against the background of direction-aware graph representation learning, including the earlier Asymmetric Graph Neural Network (AGNN), which introduced separate incoming and outgoing node embeddings for directed graphs and used an unsupervised edge-likelihood regularizer within a supervised or semi-supervised objective rather than as a fully unsupervised clustering model (Tan et al., 2021).

1. Conceptual scope and problem setting

uAGNN is presented as an unsupervised community detection approach intended especially for heterophilic graphs, where adjacent nodes are often from different classes or communities and nodes belonging to the same community may be far apart in graph distance (Leeney et al., 15 Aug 2025). In this setting, immediate neighborhoods can be misleading, because local links do not necessarily encode community membership. The method is therefore motivated by the claim that effective community detection in such graphs requires the ability to propagate information across long ranges without destroying discriminative signal.

The framework distinguishes itself from conventional GNN-based clustering methods by targeting the limitations of repeated local aggregation. The reported difficulties are local bias, oversmoothing / dissipation, and oversquashing / bottlenecks: local message passing may fail to capture non-local community structure, repeated averaging can make embeddings converge toward uninformative states, and distant information can be compressed through narrow propagation channels before reaching later layers (Leeney et al., 15 Aug 2025). Within this perspective, uAGNN is not simply a deeper local GNN, but a model whose core design objective is stable long-range propagation.

A common point of confusion concerns the relation between uAGNN and AGNN. The earlier AGNN is an asymmetric graph representation learning method for directed graphs; it learns separate incoming and outgoing embeddings and uses graph likelihood as a structural regularizer, but it is not framed as an antisymmetric dynamical system and is not primarily an unsupervised representation learner in the self-supervised or clustering sense (Tan et al., 2021). By contrast, uAGNN uses antisymmetric matrices in the dynamical-system sense and is explicitly unsupervised, with community assignments obtained after representation learning by clustering (Leeney et al., 15 Aug 2025).

2. Dynamical-system formulation and antisymmetric propagation

The central technical idea in uAGNN is to interpret graph propagation as a continuous-time dynamical system. The reported node-wise ODE is

$$\frac{d_u(t)}{dt} = f_\theta(t, _u(t), \{_v(t)\}_{v\in\mathcal{N}_u}), \qquad t\in[0,T]$$

with initial condition

$_u(0)=_u^0.$

In the non-dissipative GNN instantiation, the ODE is written as

$\frac{d_u(t)}{dt} = \sigma(_t_u(t), \Phi(\{_v(t)\}_{v\in\mathcal{N}_u})),$

where $\sigma$ is a nonlinearity, $_t \in \mathbb{R}^{d\times d}$ is a learnable weight matrix, and $\Phi$ aggregates neighborhood information (Leeney et al., 15 Aug 2025).

The stability discussion is expressed through the Jacobian $\mathbf{J}(t)$, with three regimes distinguished in the paper: Instability when $\max_i \Re(\lambda_i) > 0$, Dissipativity when $\max_i \Re(\lambda_i) < 0$, and Non-dissipativity when $\Re(\lambda_i)=0$ (Leeney et al., 15 Aug 2025). For the community-detection task considered, the desired regime is the last one, because it enables propagation that is stable without erasing information.

This motivates the antisymmetry constraint

$_u(0)=_u^0.$0

Under this condition, the paper states that the ODE becomes non-dissipative: the Jacobian has purely imaginary eigenvalues when $_u(0)=_u^0.$1 excludes the node’s own state, or eigenvalues in a small neighborhood of the imaginary axis when $_u(0)=_u^0.$2 includes self-information (Leeney et al., 15 Aug 2025). The intended effect is that the dynamics are rotational or conservative rather than contractive, so information can persist through many layers instead of collapsing toward a fixed, overly smooth state.

This design sharply contrasts with AGNN’s use of asymmetry. In AGNN, asymmetry refers to the directional distinction between sending and receiving roles in a directed graph; each node has an incoming embedding $_u(0)=_u^0.$3 and an outgoing embedding $_u(0)=_u^0.$4, and message passing proceeds in two separate streams for predecessors and successors (Tan et al., 2021). The two methods therefore address different forms of non-symmetry: uAGNN uses antisymmetric operators to shape propagation dynamics, whereas AGNN uses asymmetric representation roles to model edge directionality.

3. Discretization, layer update, and computational characteristics

Because the continuous-time system is not solved analytically in practice, uAGNN uses forward Euler discretization. The reported layerwise update is

$_u(0)=_u^0.$5

where $_u(0)=_u^0.$6 is the step size and $_u(0)=_u^0.$7 is a stability hyperparameter (Leeney et al., 15 Aug 2025). The term $_u(0)=_u^0.$8 is included to preserve numerical stability of the Euler method, while $_u(0)=_u^0.$9 acts as the antisymmetric propagation operator.

The role of $\frac{d_u(t)}{dt} = \sigma(_t_u(t), \Phi(\{_v(t)\}_{v\in\mathcal{N}_u})),$0 is described specifically as adding $\frac{d_u(t)}{dt} = \sigma(_t_u(t), \Phi(\{_v(t)\}_{v\in\mathcal{N}_u})),$1 for stability of the discretization. This controls how strongly the discrete-time system damps numerical drift while still retaining the non-dissipative behavior induced by antisymmetry (Leeney et al., 15 Aug 2025). A plausible implication is that the method seeks a balance between preserving information theoretically and keeping the finite-step numerical scheme stable in practice.

The authors further emphasize that the computational complexity is dominated by the aggregation function $\frac{d_u(t)}{dt} = \sigma(_t_u(t), \Phi(\{_v(t)\}_{v\in\mathcal{N}_u})),$2, so the antisymmetric constraint adds little overhead compared with ordinary MPNN-style propagation (Leeney et al., 15 Aug 2025). This matters because the principal architectural modification is not a wholesale replacement of neighborhood aggregation, but a structural constraint on the propagation operator combined with a dynamical-system interpretation.

The paper compares two aggregation choices: a local MPNN-like aggregator and a GCN-style normalized adjacency aggregator (Leeney et al., 15 Aug 2025). This indicates that the non-dissipative mechanism is not tied to a single neighborhood operator. Rather, the antisymmetric formulation acts as an organizing principle for deep propagation independent of the exact local aggregation rule.

4. Autoencoder architecture and unsupervised objective

uAGNN is described as an unsupervised autoencoder-style pipeline built on top of the non-dissipative GNN encoder (Leeney et al., 15 Aug 2025). The reported workflow consists of four stages: an input graph $\frac{d_u(t)}{dt} = \sigma(_t_u(t), \Phi(\{_v(t)\}_{v\in\mathcal{N}_u})),$3 with node features and adjacency; a deep non-dissipative GNN encoder that computes latent node representations $\frac{d_u(t)}{dt} = \sigma(_t_u(t), \Phi(\{_v(t)\}_{v\in\mathcal{N}_u})),$4; two decoders, one for feature reconstruction and one for adjacency reconstruction; and finally K-Means applied to the final embeddings $\frac{d_u(t)}{dt} = \sigma(_t_u(t), \Phi(\{_v(t)\}_{v\in\mathcal{N}_u})),$5 to obtain communities.

The decoding stage uses two learned linear transformations:

$\frac{d_u(t)}{dt} = \sigma(_t_u(t), \Phi(\{_v(t)\}_{v\in\mathcal{N}_u})),$6

These correspond respectively to reconstructed features and reconstructed connectivity (Leeney et al., 15 Aug 2025). The framework explicitly reconstructs both the feature matrix and adjacency matrix, in contrast to prior autoencoder methods that reconstruct only one of them.

The unsupervised training objective is

$\frac{d_u(t)}{dt} = \sigma(_t_u(t), \Phi(\{_v(t)\}_{v\in\mathcal{N}_u})),$7

Thus the model is optimized end-to-end using only feature reconstruction error and adjacency reconstruction error, with no community labels used during training (Leeney et al., 15 Aug 2025). After training, node embeddings are clustered via

$\frac{d_u(t)}{dt} = \sigma(_t_u(t), \Phi(\{_v(t)\}_{v\in\mathcal{N}_u})),$8

The resulting community detection procedure is therefore representation learning followed by clustering, rather than direct optimization of cluster assignments within the GNN. This suggests that the latent geometry learned by the non-dissipative encoder is intended to make downstream K-Means more faithful to the underlying community structure.

The contrast with AGNN is instructive. AGNN also contains an unsupervised structural component, but it is an edge log-likelihood regularizer embedded inside a supervised or semi-supervised objective (Tan et al., 2021). For a directed edge $\frac{d_u(t)}{dt} = \sigma(_t_u(t), \Phi(\{_v(t)\}_{v\in\mathcal{N}_u})),$9, AGNN models the edge probability as

$\sigma$0

using the outgoing embedding of the source and the incoming embedding of the target, and then adds the negative log-likelihood of the graph as a regularizer (Tan et al., 2021). In that model, the unsupervised signal regularizes task learning; in uAGNN, the entire objective is unsupervised reconstruction followed by clustering.

5. Experimental protocol and empirical behavior

The reported training protocol for uAGNN creates train/validation/test splits by removing edges, with 64% of edges used for training, 16% for validation, and 20% for test (Leeney et al., 15 Aug 2025). Performance is averaged across 10 random seeds. The optimizer is Adam, hyperparameters are tuned using MOTPE, and K-Means is initialized 20 times (Leeney et al., 15 Aug 2025). The main hyperparameters searched include number of layers $\sigma$1, hidden dimension $\sigma$2, $\sigma$3, $\sigma$4, and the aggregation function $\sigma$5.

The method is evaluated on ten attributed community detection datasets: Texas, Wisconsin, Cornell, EAT, BAT, UAT, CiteSeer, DBLP, Cora, and AMAP (Leeney et al., 15 Aug 2025). Their homophily values are reported as 0.11, 0.20, 0.31, 0.40, 0.45, 0.70, 0.74, 0.80, 0.81, and 0.83 respectively. Baselines include DAEGC, DGI, DMoN, GRACE, MVGRL, SUBLIME, BGRL, VGAER (Leeney et al., 15 Aug 2025).

The principal empirical pattern stated in the paper is that uAGNN performs best or near-best on low- and medium-homophily graphs, where long-range propagation matters most (Leeney et al., 15 Aug 2025). For F1, uAGNN is especially strong on Wisconsin, where it is best, and on Texas, where it is second-best; it remains competitive on BAT, EAT, and UAT. For NMI, the same trend is reported: strong performance in low- and medium-homophily settings up to about $\sigma$6, competitive behavior at high homophily, but no dominance when local structure is already highly informative (Leeney et al., 15 Aug 2025). For Conductance, where lower is better, uAGNN generally forms coherent communities by reducing cross-community edges, especially in heterophilic cases.

The dataset-specific narrative follows the same pattern. Texas is described as a prototypical low-homophily case where local GNNs struggle, and uAGNN gives strong F1 and NMI relative to baselines; Wisconsin is one of the clearest wins, with the best F1 and best NMI among compared methods; Cornell is competitive; BAT and EAT show the benefit of balancing local and non-local information; UAT remains competitive at $\sigma$7; CiteSeer, DBLP, and Cora are mixed but competitive; and on AMAP with $\sigma$8, uAGNN is not the best overall (Leeney et al., 15 Aug 2025).

These results are interpreted in the paper as evidence that the method is especially suited to non-local communities. A plausible implication is that the non-dissipative design is most useful precisely when community evidence cannot be recovered by short-range similarity alone.

6. Relation to AGNN and to the terminology of asymmetry

The naming of uAGNN can be understood more clearly by contrasting it with AGNN. AGNN, introduced as Asymmetric Graph Neural Network, addresses directed graphs where information can pass in one direction but not necessarily in the reverse direction (Tan et al., 2021). Its main architectural innovation is to assign every node two embeddings: an incoming embedding $\sigma$9, which captures its receiving role, and an outgoing embedding $_t \in \mathbb{R}^{d\times d}$0, which captures its sending role. The update rules are separated accordingly:

$_t \in \mathbb{R}^{d\times d}$1

$_t \in \mathbb{R}^{d\times d}$2

$_t \in \mathbb{R}^{d\times d}$3

$_t \in \mathbb{R}^{d\times d}$4

In its mean-pooling directed GCN-style form, AGNN uses

$_t \in \mathbb{R}^{d\times d}$5

with normalized directed adjacency matrices with self-loops (Tan et al., 2021). For node classification, the incoming and outgoing embeddings are merged, and for graph-level prediction they are pooled after node-level combination.

The key unsupervised component in AGNN is the graph likelihood term. Assuming independent Bernoulli edges, AGNN defines the directed edge probability by the source’s outgoing embedding and the target’s incoming embedding, and uses the negative graph log-likelihood as regularization (Tan et al., 2021). The total loss is

$_t \in \mathbb{R}^{d\times d}$6

where $_t \in \mathbb{R}^{d\times d}$7 is the task loss and $_t \in \mathbb{R}^{d\times d}$8 is the regularization coefficient. The regularizer is argued to help alleviate over-smoothing by preserving directed locality.

This comparison clarifies a terminological issue explicitly noted in the source material. AGNN is asymmetric in the sense of direction-aware representation learning for directed graphs; it is not antisymmetric in the algebraic or dynamical-systems sense (Tan et al., 2021). uAGNN, by contrast, is antisymmetric because it constrains the propagation operator by $_t \in \mathbb{R}^{d\times d}$9 and uses this to induce non-dissipative dynamics (Leeney et al., 15 Aug 2025). The two methods therefore share a concern with non-symmetric graph structure, but they do so through different mathematical mechanisms and for different tasks.

7. Contributions, limitations, and interpretive cautions

The paper on uAGNN claims three main contributions: the introduction of uAGNN as a novel unsupervised GNN for community detection that preserves long-range dependencies; an empirical demonstration that heterophily matters in unsupervised community detection, not only in supervised node classification; and extensive experiments showing that non-dissipative propagation improves performance in low- and medium-homophily graphs (Leeney et al., 15 Aug 2025). It also claims, to the authors’ knowledge, that this is the first work to study heterophily in unsupervised community detection in this way.

The reported limitations are equally important. uAGNN is not uniformly best: performance drops relative to some baselines on high-homophily datasets such as AMAP (Leeney et al., 15 Aug 2025). The benefit of antisymmetric long-range propagation appears smaller when the graph is already highly local and homophilic. The paper summary also notes that no explicit ablation table is provided in the excerpt, so the isolated effects of antisymmetric versus non-antisymmetric weights, of $\Phi$0, of $\Phi$1, or of reconstructing only adjacency versus only features are not directly tabulated there (Leeney et al., 15 Aug 2025). This suggests that the central explanatory claim—that the main driver is stable, non-dissipative information flow enabled by antisymmetry—rests on the model design and comparative empirical behavior rather than on a full ablation breakdown in the summarized material.

A further caution concerns the meaning of “unsupervised” across related methods. In AGNN, the unsupervised term is a structural regularizer that complements supervised or semi-supervised prediction (Tan et al., 2021). In uAGNN, unsupervised learning defines the whole training objective through reconstruction losses followed by K-Means clustering (Leeney et al., 15 Aug 2025). Conflating these settings can obscure a substantive methodological distinction.

Taken together, these works delineate two separate lines of development in graph representation learning. One line, exemplified by AGNN, adapts GNNs to directed asymmetric structure by splitting node roles into sending and receiving channels and regularizing with edge likelihood (Tan et al., 2021). The other, exemplified by uAGNN, addresses non-local community detection in heterophilic graphs by using antisymmetric propagation to maintain stable long-range signal and by learning embeddings through unsupervised reconstruction before clustering (Leeney et al., 15 Aug 2025).