acceptodds
Under review as a conference paper at ICLR 2027

Beyond Observed Topology: Semantic Graph Inference and Alignment for Text-Attributed Graph Learning

29 views

Abstract

Text-attributed graph (TAG) learning jointly models textual semantics and graph topology, typically assuming that observed edges reflect the underlying semantic structure. However, this assumption may not always hold, as connected nodes can be semantically unrelated, while semantically related nodes may lack observed connections, resulting in structural–semantic mismatch. Recent studies have explored incorporating textual semantics into graph structure, either by building upon the observed topology or by using LLMs to infer semantic relations at the edge level. While the latter offers greater flexibility, it incurs additional inference costs and limits interpretability.To address these limitations, we propose SEGIA, a SEmantic Graph Inference and Alignment approach for text-attributed graph learning. Specifically, we first introduce an uncertainty-aware semantic encoding mechanism inspired by the stochastic block model to uncover latent connectivity patterns among nodes from their textual information, yielding a semantic graph.Then, we separately encode the observed and semantic graphs to characterize their respective structural information and derive view-specific predictions. Finally, we employ optimal transport to align these predictive distributions, enabling effective information integration across the two views while enforcing prediction consistency. Extensive experiments on five public TAG benchmarks demonstrate that our proposed SEGIA outperforms state of-the-art baselines.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Text-Attributed Graphs, Graph Representation Learning, Graph Neural Network

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_RH3ZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the mismatch between observed graph structure and textual semantics in text-attributed graphs. The authors propose SEGIA, which first uses an LLM to augment node text with predicted classes and explanations, then applies a variational encoder to learn latent semantic representations, posterior uncertainty, and soft block memberships. These representations are used to construct a semantic graph through an SBM-inspired edge scorer that combines block compatibility, node activity, membership ambiguity, posterior dispersion, and latent distance. The observed and semantic graphs are then processed by separate GCN branches, whose predictions are aligned using an optimal transport objective. Experiments on five citation-network benchmarks show that SEGIA consistently outperforms the baselines considered in the paper.

Critical strengths
  1. The paper identifies a practically relevant structural–semantic mismatch in text-attributed graphs and provides a coherent framework that preserves the observed topology while introducing an additional semantic graph inferred from textual information.
  2. The overall method is clearly presented and reasonably well organized. The proposed semantic graph construction combines multiple sources of information, including soft block compatibility, node activity, semantic proximity, and posterior dispersion.
  3. The results are strong within the evaluated setting. SEGIA achieves the best reported accuracy among the compared methods on all five benchmarks, and the paper provides ablations covering the semantic graph, observed topology, OT alignment, uncertainty, SBM scoring, and LLM expansion.
Critical weaknesses
  1. The methodological novelty seems relatively incremental. Most components of SEGIA are based on established techniques, and the main contribution appears to lie in their combination rather than a fundamentally new approach.
  2. Several closely related methods discussed in the paper, especially TAPE, GAugLLM, and LLaTA, are missing from the main comparison. This makes it difficult to judge SEGIA against the most relevant alternatives.
  3. The current ablations do not isolate the individual components of the proposed edge scorer. Replacing the full scorer with cosine similarity cannot determine which terms are responsible for the improvement.
  4. The evaluation is limited to citation networks, which leaves the broader applicability to other types of text-attributed graphs unclear.
Questions for authors
  1. Could the authors provide statistics or qualitative analysis of the learned semantic graph, such as the fraction of candidate pairs not present in the observed graph, the distribution of weights assigned to observed versus newly introduced candidate pairs, or representative examples of highly and weakly weighted connections?
  2. How does SEGIA behave as the mismatch between topology and semantics increases? For example, could the authors evaluate the model under controlled edge deletion or addition?
  3. What is the computational cost of SEGIA, including LLM preprocessing and semantic graph construction, compared with the closest LLM-based TAG methods? In particular, how does the cost on OGBN-ArXiv?
3 comments

Q1: 1) Following your suggestion, we measured the proportion of candidate edges absent from the observed graph: 76.60%, 65.68%, 77.13%, 91.70%, and 24.80% for Cora, Citeseer, PubMed, ArXiv-2023, and OGBN-ArXiv, respectively. Table 1 reports the weight distributions for observed and added edges, and Table 2 reports label homophily.

Table 1. Weight distributions of observed and added edges.

Weight statisticCoraCiteseerPubMedArXiv-2023OGBN-ArXiv
Observed edges: median weight1.638±0.0900.871±0.0141.584±0.0151.154±0.0091.234±0.005
Added edges: median weight0.843±0.0220.861±0.0110.797±0.0040.590±0.0050.631±0.003
Observed edges: 90th-percentile weight1.759±0.0391.743±0.0281.641±0.0041.224±0.0111.331±0.007
Added edges: 90th-percentile weight1.694±0.0471.723±0.0251.594±0.0101.190±0.0101.164±0.008

Weights are the final symmetrized values before degree normalization.

Table 2. Label homophily (%) of observed and semantic graphs and high-weight added edges.

Graph or edge subsetCoraCiteseerPubMedArXiv-2023OGBN-ArXiv
Observed graph81.0074.3780.2464.4365.42
Semantic graph (unweighted)74.0969.2077.5362.3064.27
Semantic graph (weighted)81.41±0.5075.35±0.4383.92±0.1966.69±0.1567.89±0.05
Added edges in the top 10% by weight92.03±2.0982.59±0.9490.45±2.5876.00±0.4780.65±0.13

As shown in Table 2, the weighted semantic graph and high-weight added edges have higher label homophily than the observed graph on all five datasets. These results indicate that the learned weights emphasize class-consistent relations.

2) We also inspected representative high- and low-weight connections in Cora. Two papers on reinforcement learning for robotics (Cora IDs 55968 and 1152663) have no observed citation edge but receive a semantic edge weight of 1.8149. By contrast, an observed edge between papers on learning theory and recurrent neural networks (IDs 12350 and 16451) receives a weight of 0.0502. These examples illustrate how the scorer assigns high weights to semantically related pairs without observed edges and low weights to less related observed connections.

Q2: Following your suggestion, we evaluated SEGIA under 30% random edge addition and deletion to assess robustness to spurious and missing connections, respectively. As shown in Tables 3 and 4, SEGIA achieves the highest accuracy among the evaluated methods on all five datasets under both perturbations.

Table 3. Test accuracy (%) under 30% random edge addition.

ModelCoraCiteseerPubMedArXiv-2023OGBN-ArXiv
DeloopSGNN87.01±1.6773.54±1.7288.62±0.1572.54±0.2671.13±0.20
RADE-GCN85.98±1.7173.35±3.3684.50±0.8359.61±0.8054.53±1.25
AD-GCN83.87±1.6872.01±2.5282.89±0.7369.41±0.2668.79±0.21
CGT84.65±0.9572.60±2.2383.73±0.4965.81±0.3064.92±0.23
SEGIA88.52±1.2974.76±1.6188.89±0.5077.39±0.4175.33±0.19

Table 4. Test accuracy (%) under 30% random edge deletion.

ModelCoraCiteseerPubMedArXiv-2023OGBN-ArXiv
DeloopSGNN87.53±1.3174.36±1.5789.33±0.1574.77±0.2773.39±0.30
RADE-GCN83.62±2.8774.17±3.1076.64±2.1760.06±0.5655.30±0.87
AD-GCN85.42±2.6374.20±1.6789.18±0.7776.73±0.2674.49±0.14
CGT86.13±2.0174.42±2.4887.44±0.4373.15±0.2071.58±0.21
SEGIA88.56±1.4775.67±1.4689.55±0.5478.08±0.3676.10±0.15

Q3: Following your suggestion, we report one-time text preprocessing separately from training and evaluation (Tables 5 and 6). On OGBN-ArXiv, all three methods share the same preprocessed inputs, with a common one-time preprocessing cost of 27,778.8 s. Subsequent execution, including semantic graph construction and scoring, takes 919.25±3.94 s for SEGIA, 26,238.39±19.78 s for TAPE, and 1,672.99±15.01 s for LOGIN. SEGIA therefore has the lowest total runtime among the three methods.

Table 5. One-time text preprocessing time (min).

DatasetPreprocessing time
Cora3.43
Citeseer4.56
PubMed46.76
ArXiv-2023130.40
OGBN-ArXiv462.98
Total648.13

Table 6. Training and evaluation time per run (s; mean ± sample standard deviation).

DatasetSEGIATAPELOGIN
Cora16.51±0.15203.62±0.9759.52±0.86
Citeseer21.14±0.55247.85±0.3561.69±2.23
PubMed52.17±2.071700.63±1.60266.93±8.41
ArXiv-202340.60±0.324772.75±1.85263.99±4.02
OGBN-ArXiv919.25±3.9426238.39±19.781672.99±15.01

W1: SEGIA's contribution lies in how textual semantics and uncertainty jointly inform graph inference and how the inferred graph is integrated with the observed topology. Specifically, variational encoding provides latent representations and uncertainty estimates, which the SBM-inspired scorer uses to assign continuous weights to candidate relations. Separate graph branches retain the observed and inferred relations, while prediction alignment coordinates the two views. The semantic-graph analyses and individual scorer ablations provide evidence for these design choices.

W2: Following your suggestion, we evaluated TAPE and LOGIN using the same LLM-augmented text inputs as SEGIA (Table 7). Reproducing GAugLLM requires additional auxiliary LLM outputs and configurations, and its default Mixtral 8×7B model exceeds our computational capacity. LLaTA's code is not publicly available. We therefore included LOGIN as another closely related baseline that uses LLM feedback to guide graph refinement.

Table 7. Test accuracy (%) of closely related methods.

ModelCoraCiteseerPubMedArXiv-2023OGBN-ArXiv
TAPE87.64±1.1074.51±2.3793.50±0.5478.13±0.2376.27±0.38
LOGIN86.86±2.1072.82±2.6882.26±0.6074.64±0.4972.44±0.24
SEGIA91.00±0.7181.32±1.0390.45±0.2278.20±0.2676.54±0.06

SEGIA achieves the highest accuracy on the four datasets other than PubMed, while TAPE achieves higher accuracy on PubMed (93.50% vs. 90.45%). One possible explanation lies in their different text representation strategies. PubMed contains domain-specific biomedical terminology, for which TAPE's task-specific LM fine-tuning may provide more adaptive representations. In contrast, SEGIA employs a frozen text encoder, potentially limiting its adaptation to specialized domains. This suggests a trade-off between task-specific representation adaptation and computational efficiency. Notably, across five runs on PubMed, SEGIA requires only 3,073 s, compared with 11,308 s for TAPE, achieving approximately 3.68× faster execution.

W3: Following your suggestion, we conducted ablations that individually remove block compatibility, membership entropy, node activity, posterior variance, or latent distance from the edge scorer. As shown in Table 8, removing each scoring term reduces mean accuracy on all five datasets, supporting the contribution of each term to the full scorer.

Table 8. Test accuracy (%) with individual scorer components removed.

Model variantCoraCiteseerPubMedArXiv-2023OGBN-ArXiv
Without block compatibility89.41±1.2976.05±2.2289.34±0.1578.07±0.1176.42±0.02
Without membership entropy89.70±1.4275.96±1.9589.63±0.5277.97±0.0676.40±0.04
Without node activity89.52±1.8176.68±1.8989.55±0.5477.99±0.2176.32±0.06
Without posterior variance89.70±1.7276.46±1.4889.62±0.4378.02±0.1576.36±0.11
Without latent distance89.41±1.8375.80±1.9089.45±0.4177.97±0.3376.30±0.03
Full scorer91.00±0.7181.32±1.0390.45±0.2278.20±0.2676.54±0.06

W4: Following your suggestion, we extended the evaluation to Computers and Products, two product co-purchase graphs. As shown in Table 9, SEGIA achieves the highest mean accuracy on both datasets, supporting its applicability beyond citation networks.

Table 9. Test accuracy (%) on Computers and Products using augmented text embeddings.

ModelComputersProducts
MLP69.13±0.3680.02±0.41
GCN86.86±0.2883.81±0.47
GAT87.32±0.2282.83±0.43
DeloopSGNN87.93±0.1483.83±0.34
AdaSpec87.89±0.2881.77±0.60
RADE89.03±0.1585.18±0.45
AD-GCN87.03±0.2086.07±0.39
CGT86.60±0.2883.70±0.44
SEGIA89.09±0.4087.98±0.31

Thank you for the detailed responses and additional experiments. I appreciate the authors’ efforts to address my concerns. These additions have substantially strengthened the result. I would like to increase my rating to 3. Thank you again for the responses and discussion.

Reviewer_Gv7COpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

SEGIA augments node text with LLM predictions and rationales, then learns variational semantic representations and soft block memberships. An SBM-inspired scorer constructs a weighted semantic graph over observed, two-hop, and text-neighbor candidate edges. Two GCN branches process the observed and semantic graphs, with Sinkhorn-based prediction alignment. The paper reports improved node classification on five citation benchmarks and component ablations.

Critical strengths
  • Retaining the observed graph alongside a learned semantic graph is a reasonable way to combine complementary relational information. The candidate construction and edge-score components are explicitly defined.
  • Semantic connectivity is learned from node representations without per-edge LLM queries. The sparse candidate set limits the learned scoring stage to plausible connections, providing a practical way to refine the graph from textual evidence.
Critical weaknesses
  • The evaluation does not establish performance under the standard OGBN-ArXiv protocol. Appendix F.2 uses random 60/20/20 splits rather than the official temporal split. Table 1 also reports GraphSAGE at 32.22% and GCNII at 58.36%, without sufficient implementation detail to explain these weak results. Official-split evaluation and documented baseline inputs and tuning are needed to substantiate the state-of-the-art claim.
  • The methodological distinction remains insufficiently established. LLM-generated predictions and explanations, and text-guided graph refinement, are already central to TAPE, GAugLLM, and LLaTA, yet these methods are absent from Table 1. Comparisons under matched textual inputs are needed to establish the contribution of SEGIA's SBM-inspired construction and alignment beyond these existing ideas.
  • Evidence for the added complexity is uneven. On OGBN-ArXiv, removing OT, uncertainty, or SBM scoring costs only 0.14, 0.16, or 0.18 percentage points, with no ablation variability reported. Simpler consistency losses and a variance-penalty-specific ablation are missing. Stronger gains on other datasets do not resolve these mechanism-specific questions.
Questions for authors

Please see the weaknesses above.

1 comment

W1: 1) Following your suggestion, we evaluated the methods on the official OGBN-ArXiv temporal split. As shown in Table 1, SEGIA achieves the highest accuracy among the evaluated methods.

Table 1. Accuracy (%) on the standard OGBN-ArXiv temporal split.

ModelTest acc.
MLP69.26±0.23
GCN71.31±0.11
GAT70.68±0.05
DeloopSGNN71.88±0.42
AdaSpec71.88±0.18
RADE72.86±0.19
AD-GCN73.98±0.09
LLM4NG72.26±0.10
CGT71.22±0.18
SEGIA74.94±0.17

2) Both arXiv datasets contain 40 classes, with observed-graph homophily of 64.43% (ArXiv-2023) and 65.42% (OGBN-ArXiv). GraphSAGE's mean aggregation does not distinguish the class relevance of individual neighbors and can mix information across classes; GCNII preserves initial features, but deeper propagation can still accumulate cross-class interference. These properties may contribute to the performance differences observed on the two graphs.

W2: 1) Following your suggestion, we evaluated TAPE and LOGIN using the same LLM-augmented text inputs as SEGIA (Table 2). Reproducing GAugLLM requires additional auxiliary LLM outputs and configurations, and its default Mixtral 8×7B model exceeds our computational capacity. LLaTA's code is not publicly available. We therefore included LOGIN as another closely related baseline that uses LLM feedback to guide graph refinement.

Table 2. Accuracy (%) of closely related methods.

ModelCoraCiteseerPubMedOGBN-ArXiv (6:2:2)ArXiv-2023OGBN-ArXiv (standard)
TAPE87.64±1.1074.51±2.3793.50±0.5476.27±0.3878.13±0.2374.40±0.31
LOGIN86.86±2.1072.82±2.6882.26±0.6072.44±0.2474.64±0.4972.16±0.24
SEGIA91.00±0.7181.32±1.0390.45±0.2276.54±0.0678.20±0.2674.94±0.17

As shown in Table 2, SEGIA achieves the highest accuracy in the other five reported settings, while TAPE is higher on PubMed (93.50% vs. 90.45%). One possible explanation is their text representation strategies: PubMed's domain-specific terminology may suit TAPE's task-specific LM fine-tuning, whereas SEGIA's frozen encoder may limit adaptation to specialized domains—a trade-off between representation adaptation and computational efficiency. Over five runs on PubMed, SEGIA requires 3,073 s versus 11,308 s for TAPE (both including preprocessing), approximately 3.68× faster.

2) We evaluated competitive baselines using the same augmented text embeddings as SEGIA. As shown in Table 3, SEGIA achieves the highest accuracy under matched-input conditions.

Table 3. Accuracy (%) with identical augmented embeddings.

ModelCoraCiteseerPubMedOGBN-ArXivArXiv-2023
MLP83.65±1.3674.26±2.1988.49±0.5569.52±0.1273.60±0.60
GCN88.93±1.6874.76±1.9788.10±0.5272.62±0.2373.83±0.44
GAT88.52±1.7574.33±2.3986.51±0.4671.72±0.2570.52±0.36
DeloopSGNN88.60±2.0474.98±1.8688.99±0.4473.47±0.3274.55±0.43
AdaSpec88.75±2.1473.76±2.0888.42±0.5373.01±0.2772.13±0.33
RADE88.12±2.5172.76±2.0888.68±0.6474.59±0.3775.06±0.34
AD-GCN86.83±1.9275.05±1.2989.62±0.6274.86±0.2376.43±0.19
CGT88.38±2.2675.89±1.7087.66±0.3273.46±0.3072.48±0.28
SEGIA91.00±0.7181.32±1.0390.45±0.2276.54±0.0678.20±0.26

W3: 1) Following your suggestion, we report repeated-run results on OGBN-ArXiv for the variants without OT, uncertainty, and SBM scoring. Their mean accuracies are lower than that of the full model, with standard deviations of 0.02, 0.08, and 0.07 percentage points, respectively.

2) We evaluated two ablations: removing only the variance penalty while retaining posterior sampling, and replacing posterior samples with the posterior mean while retaining the predicted variance and its scoring penalty. All other settings were held fixed (Table 4).

Table 4. Accuracy (%) with ablations of the variance penalty and posterior sampling.

Model variantCoraCiteseerPubMedArXiv-2023OGBN-ArXiv
Without variance penalty89.67±1.5979.94±1.2589.55±0.4977.43±0.3576.22±0.09
Deterministic posterior mean88.97±2.0679.55±1.1189.50±0.4777.39±0.4876.28±0.10
Full model91.00±0.7181.32±1.0390.45±0.2278.20±0.2676.54±0.06

3) Following your suggestion, we replaced Sinkhorn alignment with JS divergence or total variation (TV), keeping all other settings fixed. As shown in Table 5, Sinkhorn alignment achieves higher mean accuracy on all five datasets.

Table 5. Accuracy (%) with different alignment losses.

Alignment lossCoraCiteseerPubMedArXiv-2023OGBN-ArXiv
JS divergence89.30±2.3879.84±1.5889.66±0.5877.31±0.2976.24±0.11
TV89.89±1.3180.36±1.5589.93±0.5277.33±0.1976.20±0.14
Sinkhorn91.00±0.7181.32±1.0390.45±0.2278.20±0.2676.54±0.06
Reviewer_uBr1OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SEGIA to deal with the mismatch between the observed topology and the text semantics in text-attributed graphs. The method first uses an LLM to generate class predictions and explanations, concatenates them with the raw text, and obtains node features with a frozen text encoder. Then, a variational encoder produces latent representations, posterior variances and soft cluster assignments. Candidate edges are scored by combining block compatibility, node activity, membership entropy, posterior variance and latent distance, which gives a weighted semantic graph. Two independent GCNs process the observed graph and the semantic graph respectively. They are trained jointly with classification supervision and a Sinkhorn-based prediction alignment loss, and the final prediction is the average of the logits of the two branches.

The authors report node classification results on five citation networks, together with component ablations and hyperparameter sensitivity analysis. However, the authors' own ablation results show that the main advantage over the baselines already exists without the semantic graph, and the core contribution of the paper, i.e., the semantic graph inference and its related mechanisms, is not directly supported by evidence.

Critical strengths
  1. The research problem is practically meaningful. Observed edges do not always fully reflect the task-related textual relations, so introducing an additional text-based semantic graph is well motivated. SEGIA keeps the observed graph and models the semantic graph with an independent branch, which is a design direction worth studying.

  2. The inference cost is relatively controllable. Compared with methods that call the LLM for each edge, SEGIA only calls the LLM once for each node, and the edge scoring is done by a lightweight model.

  3. The paper provides empirical results on several datasets. The experiments cover five citation networks, report the mean and standard deviation over five runs, and include ablations on the branches, the candidate edge sources and several components. The full prompts are also given in the appendix.

Critical weaknesses
  1. The comparisons do not sufficiently isolate the contribution of semantic graph inference. The paper does not clearly specify whether the baselines in Table 1 use the same LLM-augmented text, text encoder, and comparable tuning budgets. Closely related methods such as TAPE, GAugLLM, and LLaTA are also absent from the comparison. Table 2 shows that the variant without the semantic graph already outperforms every baseline in Table 1 on ArXiv-2023 (75.07%) and OGBN-ArXiv (75.92%). On PubMed, much of the gap between GCN (50.11%) and SEGIA (90.45%) is already present in this variant (87.88%). These observations do not negate the improvements from adding the semantic graph, which Table 2 does demonstrate, but they leave unclear how much of the advantage over existing methods arises from graph inference versus other differences in inputs, architecture, or training. Matched comparisons are therefore needed to substantiate the attribution of the reported gains. Reproducibility is further limited by Appendix F.3 providing only a reference configuration rather than the actual per-dataset configurations. Finally, the random 6:2:2 split limits direct comparison with results under the official OGBN-ArXiv temporal protocol, although it does not by itself invalidate comparisons conducted under identical splits.

  2. The quality of the semantic graph and the core mechanisms are not verified. There is no direct analysis of the semantic graph, such as the homophily of added or down-weighted edges. The indirect evidence is negative: using only the semantic graph on Cora (79.15%) is worse than MLP (84.13%). The "w/o Uncertainty" variant only replaces sampling with the posterior mean, so it does not test the variance penalty in Eq. (21). For OT, the cost matrix is symmetric, so the bidirectional average in Eq. (31) is redundant, and without entropic regularization the cost reduces to total variation distance, no comparison with such simple losses is given. Also, the edge scorer is supervised by observed edges, which makes it tend to reproduce the observed topology rather than correct it.

  3. The evaluation does not match the motivation. All five datasets are homophilic citation networks, where the structure-semantics mismatch is the least serious, and no heterophilic text-attributed graph is tested.

Questions for authors
  1. What text features and LLM augmentation does each baseline use? Can the authors provide matched comparisons (e.g., GCN/TAPE with the same augmented text and tuning budget), together with the per-dataset configurations? Given that the variant without the semantic graph already exceeds all baselines, how do the authors argue that the semantic graph is the main source of the improvement?

  2. Can the authors report the label homophily of the semantic graph, and of the added and down-weighted edges, compared with the observed graph? What evidence shows that the scorer corrects the semantic mismatch rather than just fitting the observed topology?

  3. Does "w/o Uncertainty" remove the variance penalty in Eq. (21)? Can the authors separately ablate the variance penalty, and compare the Sinkhorn alignment with simpler consistency losses such as total variation or JS divergence?

2 comments

Q1&W1: 1) The baselines in Table 1 of the paper were run with each method's default configuration, using the original dataset features and without LLM augmentation. We additionally evaluated competitive baselines with the same augmented text embeddings as SEGIA (Table 1), where SEGIA achieves the highest accuracy under matched-input conditions.

Table 1. Test accuracy (%) using identical augmented text embeddings.

ModelCoraCiteseerPubMedArXiv-2023OGBN-ArXiv
MLP83.65±1.3674.26±2.1988.49±0.5573.60±0.6069.52±0.12
GCN88.93±1.6874.76±1.9788.10±0.5273.83±0.4472.62±0.23
GAT88.52±1.7574.33±2.3986.51±0.4670.52±0.3671.72±0.25
DeloopSGNN88.60±2.0474.98±1.8688.99±0.4474.55±0.4373.47±0.32
AdaSpec88.75±2.1473.76±2.0888.42±0.5372.13±0.3373.01±0.27
RADE88.12±2.5172.76±2.0888.68±0.6475.06±0.3474.59±0.37
AD-GCN86.83±1.9275.05±1.2989.62±0.6276.43±0.1974.86±0.23
CGT88.38±2.2675.89±1.7087.66±0.3272.48±0.2873.46±0.30
SEGIA91.00±0.7181.32±1.0390.45±0.2278.20±0.2676.54±0.06

Under matched inputs, some baselines outperform w/o Semantic, while SEGIA achieves the highest accuracy (Table 1). These results support the additional contribution of semantic graph inference.

2) Our SEGIA runs use the same 6:2:2 split and 200 epochs. Common weights are KL = 0.01, reconstruction = 0.1, view classification = 0.5, graph OT = 0.5, SBM α = 1.0, structural coefficient = 0.5, and SBM γ = 0.5.

ParameterCoraCiteseerPubMedArXiv-2023OGBN-ArXiv
edge-BCE0.51.00.50.50.5
SBM activity0.50.51.01.00.5
semantic top-k535101
two-hop top-k5125101
negative-edge ratio1.01.01.01.00.01
Sinkhorn iter.2050202010

3) Following your suggestion, we evaluated TAPE and LOGIN using the same LLM-augmented text inputs as SEGIA (Table 2). Reproducing GAugLLM requires additional auxiliary LLM outputs and configurations, and its default Mixtral 8×7B model exceeds our computational capacity. LLaTA's code is not publicly available. We therefore included LOGIN as another closely related baseline that uses LLM feedback to guide graph refinement.

Table 2. Test accuracy (%) of closely related methods.

ModelCoraCiteseerPubMedArXiv-2023OGBN-ArXiv
TAPE87.64±1.1074.51±2.3793.50±0.5478.13±0.2376.27±0.38
LOGIN86.86±2.1072.82±2.6882.26±0.6074.64±0.4972.44±0.24
SEGIA91.00±0.7181.32±1.0390.45±0.2278.20±0.2676.54±0.06

SEGIA achieves the highest accuracy on the four datasets other than PubMed, while TAPE achieves higher accuracy on PubMed (93.50% vs. 90.45%). One possible explanation is their different text representation strategies: PubMed's domain-specific terminology may suit TAPE's task-specific LM fine-tuning, whereas SEGIA's frozen encoder may limit adaptation to specialized domains, a trade-off against computational efficiency. Notably, across five runs on PubMed, SEGIA requires only 3,073 s versus 11,308 s for TAPE (both with one-time preprocessing), about 3.68× faster.

Q2&W2: 1) Label homophily and edge-weighting mechanism. As recommended, we report label-homophily measurements for the observed graph, the semantic graph, newly added edges, and observed edges with relatively low weights (Table 3), assessing whether learned weights favor same-class connections.

Table 3. Label homophily (%) of observed and semantic graphs and edge subsets.

Graph or edge subsetCoraCiteseerPubMedArXiv-2023OGBN-ArXiv
Observed graph81.0074.3780.2464.4365.42
Semantic graph (unweighted)74.0969.2077.5362.3064.27
Semantic graph (weighted)81.41±0.5075.35±0.4383.92±0.1966.69±0.1567.89±0.05
All added edges (unweighted)71.9866.5076.7362.1160.78
All added edges (weighted)79.18±0.5973.31±0.6183.73±0.2266.62±0.1664.80±0.03
Added edges among the top 10% of all edges by weight92.03±2.0982.59±0.9490.45±2.5876.00±0.4780.65±0.13
Observed edges with the lowest weights (bottom 25%)55.13±1.8352.16±2.3758.35±3.8144.12±1.0642.28±1.16

2) Observed-edge supervision and semantic-graph expansion. Observed-edge supervision is not equivalent to reconstructing the observed adjacency matrix. Observed edges, text-neighbor edges, and two-hop edges are evaluated by a shared continuous scoring function that incorporates latent representations, soft block memberships, uncertainty, and semantic distance. The inferred graph is therefore not restricted to the observed edge set. As shown in Table 3, the weighted semantic graph and high-weight added edges have higher homophily than the observed graph on all five datasets. These results indicate that the scorer assigns high weights to class-consistent relations beyond the observed graph.

Q3&W2: 1) Variance-penalty and posterior-sampling ablations. The w/o Uncertainty implementation jointly disables posterior sampling, variance statistics, and the variance penalty in Equation (21). To distinguish the effects of posterior sampling and the variance penalty, we evaluated two ablations: removing only the variance penalty while retaining posterior sampling, and replacing posterior samples with the posterior mean while retaining the predicted variance and its scoring penalty. All other settings were held fixed (Table 4).

Table 4. Test accuracy (%) with ablations of the variance penalty and posterior sampling.

Model variantCoraCiteseerPubMedArXiv-2023OGBN-ArXiv
Without variance penalty89.67±1.5979.94±1.2589.55±0.4977.43±0.3576.22±0.09
Deterministic posterior mean88.97±2.0679.55±1.1189.50±0.4777.39±0.4876.28±0.10
SEGIA91.00±0.7181.32±1.0390.45±0.2278.20±0.2676.54±0.06

2) Alignment-loss comparison. We replaced Sinkhorn alignment with Jensen–Shannon (JS) divergence or total variation (TV), keeping all other settings fixed. As shown in Table 5, Sinkhorn alignment achieves higher mean accuracy on all five datasets.

Our implementation obtains an entropy-regularized transport plan using Sinkhorn iterations and uses the transport cost under that plan as the alignment loss. It therefore differs from the unregularized OT formulation that is equivalent to TV under the relevant cost. For a symmetric cost, the two transport directions are equivalent at exact convergence; averaging both directions symmetrizes the finite-iteration computation.

Table 5. Test accuracy (%) with different alignment losses.

Alignment lossCoraCiteseerPubMedArXiv-2023OGBN-ArXiv
JS divergence89.30±2.3879.84±1.5889.66±0.5877.31±0.2976.24±0.11
TV89.89±1.3180.36±1.5589.93±0.5277.33±0.1976.20±0.14
SEGIA91.00±0.7181.32±1.0390.45±0.2278.20±0.2676.54±0.06

W1: Official temporal split. Regarding your concern that “the random 6:2:2 split limits direct comparison with results under the official OGBN-ArXiv temporal protocol,” we evaluated the methods on the official temporal split. As shown in Table 6, SEGIA achieves the highest accuracy among the evaluated methods.

Table 6. Test accuracy (%) on the official OGBN-ArXiv temporal split.

ModelTest accuracy
MLP69.26±0.23
GCN71.31±0.11
GAT70.68±0.05
DeloopSGNN71.88±0.42
AdaSpec71.88±0.18
RADE72.86±0.19
AD-GCN73.98±0.09
LLM4NG72.26±0.10
CGT71.22±0.18
TAPE74.40±0.31
LOGIN72.16±0.24
SEGIA74.94±0.17

W2: Observed-topology ablation. Regarding your observation that “using only the semantic graph on Cora (79.15%) is worse than MLP (84.13%),” the w/o Observed Topology variant removes the observed branch, observed/two-hop candidates, and cross-view alignment. It discards useful structure in Cora (observed-edge homophily 81.00%), while semantic-neighbor aggregation may mix classes. The full model retains both relational sources.

W3: Edge-perturbation robustness. To examine your concern that “The evaluation does not match the motivation,” we evaluated SEGIA under 30% random edge addition and deletion. As shown in Table 7, SEGIA achieves the highest accuracy among the evaluated methods on all five datasets under both perturbations, supporting its robustness to spurious and missing connections.

Table 7. Test accuracy (%) under 30% random edge addition/deletion (mean ± SD).

ModelCora (Add/Delete)Citeseer (Add/Delete)PubMed (Add/Delete)ArXiv-2023 (Add/Delete)OGBN-ArXiv (Add/Delete)
DeloopSGNN87.01±1.67 / 87.53±1.3173.54±1.72 / 74.36±1.5788.62±0.15 / 89.33±0.1572.54±0.26 / 74.77±0.2771.13±0.20 / 73.39±0.30
RADE-GCN85.98±1.71 / 83.62±2.8773.35±3.36 / 74.17±3.1084.50±0.83 / 76.64±2.1759.61±0.80 / 60.06±0.5654.53±1.25 / 55.30±0.87
AD-GCN83.87±1.68 / 85.42±2.6372.01±2.52 / 74.20±1.6782.89±0.73 / 89.18±0.7769.41±0.26 / 76.73±0.2668.79±0.21 / 74.49±0.14
CGT84.65±0.95 / 86.13±2.0172.60±2.23 / 74.42±2.4883.73±0.49 / 87.44±0.4365.81±0.30 / 73.15±0.2064.92±0.23 / 71.58±0.21
SEGIA88.52±1.29 / 88.56±1.4774.76±1.61 / 75.67±1.4688.89±0.50 / 89.55±0.5477.39±0.41 / 78.08±0.3675.33±0.19 / 76.10±0.15

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.