acceptodds
Under review as a conference paper at ICLR 2027

Taco-Graph: Automated Construction of Text-Attributed Graphs

Abstract

Graph-structured representations of text data have become foundational across machine learning, yet it remains unclear how to systematically construct a graph from a given dataset for a target task. Choices about which entities become nodes, which relationships become edges, and what text each node contains can substantially affect downstream performance. In practice, finding optimal constructions often requires expert knowledge or manual trial and error. We introduce TACO-Graph, a decision tree framework for learning which graph construction choices are most effective for a given dataset and task. Specifically, we focus on text-attributed graphs (TAGs) constructed from multi-attribute tabular data. TACO-Graph enumerates valid TAG representations, scores each using fast proxy evaluation on small representative subsets, and trains an interpretable decision tree on those scores. Across five datasets spanning citation networks, product reviews, and bibliographic relationships, our decision trees predict construction quality on separately drawn evaluation samples with mean Spearman above for node classification and above for GraphRAG retrieval. Our trees systematically identify which constructions perform better across both GNN and GraphRAG tasks, and the full pipeline is released as an open-source library to support evaluation on new tasks and datasets. TACO-Graph provides a systematic and interpretable guide for determining how to convert tabular data to text-attributed graphs for downstream tasks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
text-attributed graphs, graph construction, decision tree meta-learning, graph neural networks, GraphRAG, retrieval-augmented generation, interpretable machine learning, construction quality prediction

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

The paper presents TACO-Graph, a framework for selecting text-attributed graph constructions from tabular datasets. It organizes construction choices along three axes: node entities, edge relationships, and textual attributes. Valid configurations are evaluated using GraphSAGE classification or GraphRAG question answering, and a decision tree is fitted to their mean scores to derive interpretable construction rules. Experiments on five datasets show strong agreement between predicted rankings and rankings measured on independently drawn evaluation samples. The paper also studies ranking stability as the evaluation budget increases.

Critical strengths
  • Relevant problem formulation. Graph construction is often treated as a fixed preprocessing step, despite its potential impact on downstream performance. Explicitly evaluating these choices is useful for practitioners.
  • The writing of paper is clear.
Critical weaknesses
  • First of all, I think for a top-tier conference, the paper’s methodological contribution is insufficient, and the methodology is described in very limited detail.
  • Comparability of the classification task across node types is insufficiently specified. The search changes the prediction unit between primary, secondary, and aggregate entities. For example, Amazon configurations use reviews, reviewers, and product families, while the task is described as predicting product subcategories. A reviewer can be associated with multiple categories, just as an author can publish papers in multiple subject areas. The paper does not sufficiently explain how labels and text are assigned to these entities, how conflicting labels are handled, or whether predictions are evaluated against a common set of underlying test examples. Consequently, differences attributed to construction quality may also reflect changes in target definition, label distribution, text aggregation, or task difficulty.
  • The GNN experiments lack a text-only classifier using the same embeddings and a corresponding model with edges removed. These baselines would help determine how much of the observed performance variation comes from graph structure rather than textual information and entity aggregation. Replacing all embeddings with zeros, as in T6e, does not answer this question.
Questions for authors
  • In Figure 2, Sports primary-node classification scores are approximately 55–65, whereas the corresponding Figure 3 leaf values are approximately 39.5–41.3. Section 3.2.1 defines a pseudo-(R^2) score floored at zero, but Section 4.1 describes no-text controls as remaining at the majority-class floor, and Figure 2 shows a nonzero control curve. Section 3.2.2 describes question-level faithfulness filtering, whereas Appendix A describes subtype-level pilot filtering. The procedure producing History’s 100/100/33 questions is not sufficiently clear. Can authors explain these?
  • Can the authors report the agreement between proxy rankings and full-scale GNN rankings, together with the final performance of the selected configuration relative to the full-scale best?
1 comment
AuthorsOpenReview ↗

Plot, Baseline + Appendix Details

Thank you very much for the thorough review! We particularly appreciate your close reading of the numbers across our figures, which caught plotting inconsistencies we have now fixed. We respond in four sections.

  1. Inconsistencies in the Plot values

Thank you for catching the inconsistent numbers! We save both raw accuracy and the McFadden pseudo-R² for every run. Figure 2 displays accuracy while Figure 3's leaves display pseudo-R², which is why one reads 55–65 and the other 39.5–41.3. This is a labelling error on our part rather than a disagreement in the underlying data. It also explains your second observation: the no-text control sits at the majority-class floor on the accuracy scale, which is nonzero, but correctly at zero on the pseudo-R² scale (Section 3.2.1). We have put both figures on pseudo-R², updated the values quoted in Section 4.1, and made Section 3.2.1 explicit about which quantity is which. On question filtering, Section 3.2.2 and Appendix A describe the same procedure at different granularities, and we have rewritten both to describe the single procedure we run. For History's 100/100/33 split, we generate 100 questions each for the single-entity and multi-hop subtypes and 33 for the aggregate subtype, because History's aggregate entity is too coarse to support 100 distinct aggregate questions (even with multiple attempts using GPT 4o-mini).

  1. Comparability of the classification task across node types

We describe method-level components in the body and delegate dataset specifics to Appendix A. However, Appendix A currently does not specify how labels are assigned when the prediction unit changes with node type, and we have itemized this detail per dataset in this section now. For Sports, N1 predicts a review's product subcategory, N2 a reviewer's modal subcategory across their reviews, and N3 a product family's subcategory, which is deterministic since each family maps to one subcategory. We also believe the McFadden pseudo-R² at least partly addresses your label distribution concern: it measures lift over each task's own class prior, so a node type with a more skewed class distribution gets no credit for that. It does not remove differences in how predictable each target is, and we agree that choosing a node type inherently means choosing the prediction unit. We now state this in Section 3.2.1 and list a fixed-target evaluation, scoring every node type against review-level labels, as a limitation.

  1. Text Only Classifier / structure-removed baseline

You are right that T6e does not answer whether graph structure adds value beyond text, since zeroing the text removes the wrong thing. We do train a structure-blind MLP (which is saved in our anonymized code) on identical embeddings alongside every GNN run, but never reported it. Averaged over variants, the GNN's lift over the MLP is +1.4 points on History and +1.7 on ArXiv, but −1.3 on Sports, −0.9 on Electronics and −1.0 on Toys. These gaps are small next to the spread across constructions on most datasets (50 points on Electronics, 46 on ArXiv, 32 on Sports), so message passing changes scores far less than node and text choices do. The construction ranking is strongly preserved between the two models (ρ of 0.86 to 0.96), so for node classification our trees mostly rank node and text choices rather than edges, with edge type receiving only 0–2% of tree importance across all five datasets. However, edge choice matters far more for GraphRAG, where retrieval trees assign 62–100% of importance to edge type on all five datasets (Appendix B). We read this as evidence that the value of edge construction depends on how much the task requires traversing the graph, which is additional evidence towards varying value of axes depending on the task you are trying to accomplish. We have added the MLP comparison to Table 3 and Appendix E, and revised Section 4.1 accordingly.

  1. Proxy versus full-scale agreement

Yes, we have run this (per-dataset table in Appendix C, and in our responses to Reviewers JxMn and 6riB). Against full scale rankings our proxy gives a Spearman ρ of 0.75–0.90 (mean 0.82) compared to 0.97 in Table 3, all statistically significant. In selection terms, the construction we rank first is the full scale optimum on three of five datasets, with a mean top-1 regret of 0.9 points and a top-3 regret of at most 0.1 points everywhere (out of 100). The lower ρ comes mostly from reordering along the edge axis rather than at the top of the ranking, because edge rules scale differently: a construction that is sparse at 1,000 rows can be near-complete at 50,000, while other rules leave 48–76% of nodes isolated. Together, these show that subset rankings identify near-optimal full-scale constructions, and we now report both side by side in Section 4.3.

We will upload the revised and reformatted PDF on Monday 11/16.

Reviewer_6riBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to systematically construct text-attributed graphs (TAGs) from multi-attribute tabular data for a given downstream task. The authors propose TACO-Graph, which defines a discrete construction space over node entities, edge types, and text fidelity, enumerates valid graph variants, evaluates them using lightweight proxy runs on representative subsets, and fits interpretable decision trees to predict construction quality. The framework is evaluated on five datasets covering citation, e-commerce, and bibliographic data, with downstream tasks including GNN-based node classification and GraphRAG retrieval. The results show that graph construction choices can substantially affect downstream performance, and that construction rankings learned from small subsets are relatively stable on separately drawn evaluation samples, with within-dataset Spearman correlations above 0.9 for node classification and above 0.8 for GraphRAG retrieval.

Critical strengths
  1. The paper addresses a practical problem that is often overlooked in TAG learning. Graph construction choices can strongly affect downstream performance, but many existing methods assume a fixed graph. Defining node, edge, and text choices as explicit construction variables is useful and well motivated.

  2. The experimental validation is reasonable. The authors evaluate ranking consistency using Spearman correlation and tree edit distance. They also conduct additional experiments with strictly disjoint data pools to examine whether sample overlap affects the results.

Critical weaknesses
  1. The contribution of the decision-tree meta-learner is not sufficiently established. All valid graph variants are already proxy-evaluated before fitting the tree. It is unclear whether the decision tree improves construction selection compared with directly ranking variants by their mean proxy scores. The experiments also lack direct comparisons with existing automated graph construction methods.

  2. The construction space is relatively limited and still requires manual setup. For a new dataset, users must provide schema mapping, define the label space, and specify valid variants. Some edge construction methods also use fixed parameters. The sensitivity of the learned rankings to these choices is not evaluated.

  3. The robustness of construction rankings across downstream model architectures remains unclear. The node-classification experiments use a fixed 2-layer GraphSAGE model, while GraphRAG uses a single retrieval and generation pipeline. The results therefore mainly support ranking consistency under the evaluated settings.

  4. The main experimental figures could present the findings more directly. Figure 2 mainly shows performance differences across construction variants, while Figure 3 provides examples of learned decision trees. These results are useful, but they do not directly show whether the proposed selection framework performs better than simpler alternatives.

Questions for authors
  1. Since all valid graph variants are already evaluated on proxy subsets, could the authors compare direct ranking by mean proxy scores with decision-tree predictions using Spearman correlation on the test pool?

  2. Could the authors evaluate the selected constructions at full scale on at least one dataset? A comparison between the decision-tree selection, direct proxy-score ranking, and the best full-scale construction would help validate the practical benefit.

  3. How stable are the learned construction rankings across downstream model architectures? Could the authors repeat the experiments with another GNN architecture and report the ranking correlation?

  4. How sensitive are the learned decision rules to different edge construction parameters and text-fidelity choices?

2 comments
AuthorsOpenReview ↗

Decision Tree Evaluation + Stability Notes

Thank you very much for the keen review! We particularly appreciate your recognition that graph construction is a practical, often overlooked problem in TAG learning, and your notes on validating the tree against simpler alternatives and testing stability across architectures. We respond in two sections.

  1. Decision Tree Evaluation

    A.: First, we'd like to clarify a point that we realize must be made clearer in our body! The decision tree is not meant to improve selection, its purpose is for interpretability. Ranking variants directly by their mean proxy score gives a test-pool Spearman ρ of 0.98 on average, against 0.97 for the decision tree predictions, and the plain mean is marginally ahead on every dataset. This is expected and not a problem for our claim: the decision tree is not meant to improve selection, its purpose is primarily interpretability. The tree assigns one predicted value per leaf, so it deliberately smooths over small differences and noise between variants that share axis levels, and that compression is the point: the raw ranking tells you which construction is best, not why. Because our method decomposes constructions into node, edge and text choices, the tree allows you to say things like 'the edge axis does not bear much weight compared to the choice of node or text', which in a practitioner's setting can be as useful as having the most performant construction. Another thing the tree does that the raw ranking cannot is rank a construction it has never scored. We ran a leave-one-variant-out experiment where the tree is refit with one construction held out and then predicts it, giving ρ of 0.45–0.93 for node classification and 0.54–0.79 for GraphRAG. The plain mean has nothing to say about an unscored variant, so this is where the meta-learner is additionally useful. Both comparisons are now in the ablation table in Appendix C.

    B: We have now run full scale evaluations on all five datasets, exactly the comparison you describe (per-dataset table in Appendix C, and in our response to Reviewer JxMn). Against the full scale rankings our proxy gives ρ of 0.75–0.90 (mean 0.82), and in selection terms the construction we rank first is the full scale optimum on three of five datasets, with a mean top-1 regret of 0.9 points and a top-3 regret of at most 0.1 points everywhere. We also compare against the simpler alternatives in your fourth weakness: a blind pick, the default (rows as nodes with full text), and the oracle. Our pick recovers 86–100% of the oracle gain over a blind pick for node classification and 62–100% for GraphRAG, and beats the default on 4 of 5 datasets for node classification and 5 of 5 for GraphRAG. The one case where it does not is Sports (40.17 against the default's 41.00), which we think is informative: Sports' top constructions are tightly clustered (Figure 2), so the default is already near optimal. We agree these belong in a figure and have added one to Section 4.3.

  2. Stability Notes

    A.: Yes, we have now run GCN in addition to GraphSAGE! We rescored every valid M1 variant on ten train-pool subsets with GCNConv substituted for SAGEConv, and rescored GraphSAGE on the same subsets so the two are compared on matched data. Rank correlation between the two architectures is 0.76–0.94 across the five datasets, all statistically significant, with per-dataset values in Appendix E. The top-ranked node type is identical on all five datasets, and the full top-ranked construction on Electronics and Toys; where they disagree, it is always on the edge choice, never the node or text choice. This matches our trees, which already identify the edge axis as carrying the least weight.

    B.: On the sensitivity of the learned rules, our ablation experiments vary which constructions are available to the tree and report how much of the oracle gain survives (the 86–100% figure above). To be clear about what we have not done yet: we did not sweep the fixed edge parameters themselves (k=5 for structural similarity, 0.75 cosine for semantic similarity). These were calibrated and held constant across all variants and datasets, so they cannot favour one construction over another, but whether a different k would reorder the edge axis is an open question, which we have added to the limitations in Section 5. We are happy to run a small k sweep during the discussion period. Your point about manual setup is fair: a new dataset needs a schema mapping, label definitions, and a variant list. While submitting, we began to build an automated front end that infers the schema mapping from column names and shapes, derives the mechanical exclusions from measured data properties, and refuses datasets whose construction space is too thin to rank. It is still under development, and we now note it in Appendix A ('Portability to a new dataset') rather than claim it here, and would be glad to describe it further.

We will upload the revised and reformatted PDF on Monday 11/16.

Thank you for the detailed response and additional experiments. The full scale evaluation and cross-architecture results address several of my concerns. I also appreciate the clarification of the decision tree's role.

However, I still have some concerns about the novelty and the limited construction space. The sensitivity to fixed edge parameters also remains unclear. I encourage the authors to discuss these limitations more clearly in the revised paper.

Overall, the new results strengthen the paper, but I will keep my original rating.

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

Taco-Graph searches over text-attributed graph constructions from multi-attribute tabular data. Its design space varies the node entity, edge construction, and text fidelity. Valid variants are scored using small-sample GNN node classification or GraphRAG evaluation, and a shallow decision-tree regressor learns to predict construction quality and expose interpretable design rules. Experiments cover five datasets and two downstream tasks. The paper reports high within-dataset Spearman correlations between tree predictions and scores on separately drawn evaluation data, additional strictly disjoint-pool checks, cross-dataset transfer analyses, and evaluation-cost measurements. For GraphRAG, the method uses GPT-4o-mini for question generation, answering, and RAGAS judging; its whole-space evaluation costs 8.0–27.5 times one full-scale run.

Critical strengths
  1. The paper addresses an important and under-systematized design problem. Treating node identity, relational edges, and text content as explicit construction choices is useful for both TAG learning and GraphRAG. The unified framework makes decisions that are often implicit and ad hoc visible, and the observed variation across datasets and tasks convincingly shows that construction choice matters.

  2. The empirical analysis is systematic and candid about several limitations. The authors evaluate multiple datasets, check separately drawn and strictly disjoint pools, examine ranking stability as a function of evaluation budget, and release code and configuration files. I also appreciate the explicit reporting that cross-dataset GraphRAG transfer is inconsistent and that the current GraphRAG pipeline does not provide runtime savings over a small number of full-scale evaluations.

Critical weaknesses
  1. The central deployment claim is validated only at the proxy level, and the value of the decision tree is not isolated. For GNNs, both training and validation rankings come from small subsets; for GraphRAG, they come from different questions evaluated on the same pooled corpus. The paper does not show that the selected construction improves actual full-data performance over direct proxy-mean selection, random selection, or a small full-scale search. High Spearman correlation can coexist with poor top-1 or top-k selection, which is what matters in deployment. Moreover, after every variant has been scored on the proxy pool, selecting the highest proxy mean is an obvious baseline that does not require a decision tree. Without this comparison and full-scale top-k regret or realized performance, the practical benefit of the meta-learner remains unclear.

  2. The GraphRAG evidence conflates graph-dependent retrieval with generator and judge behavior. GPT-4o-mini generates the questions, produces the answers, and supplies the RAGAS judgments, with no human validation or conventional retrieval metrics reported. The composite score mixes faithfulness, answer relevance, and context relevance, making it difficult to determine whether graph construction improves retrieval itself or changes downstream generation. This concern is amplified by the substantial whole-space cost and the absence of ranking convergence at smaller question budgets.

Questions for authors
  1. Can you evaluate the selected constructions on the full datasets and compare against direct proxy-mean selection, random selection, and exhaustive/full-scale oracle choices? Reporting top-1/top-k regret and realized full-scale performance would directly test practical utility. What additional benefit does the decision tree provide over simply ranking variants by their proxy means?

  2. For GraphRAG, can you separate retrieval quality from generation quality using recall/precision of gold-supporting passages or answer accuracy, and validate a subset with human judgments or an independent judge? Does the selected graph construction remain consistent under these alternative evaluations?

2 comments
AuthorsOpenReview ↗

Decision Tree Evaluation + GraphRAG Robustness

Thank you very much for the insightful review! We particularly appreciate your recognition that graph construction is important and under-addressed, and your emphasis on whether our decision trees hold up with more data and whether the GraphRAG evaluation is robust. We respond in two sections.

  1. Decision Tree Evaluation

    A.: First, we'd like to clarify a point that we realize must be made clearer in our body! The decision tree is not meant to improve selection, its purpose is for interpretability. While you can absolutely use the pure variant scores alone to choose the best graph construction, doing so does not clarify at all why the construction is best. Because our method decomposes construction choices, the tree shows which elements contribute most to performance and smooths over small differences and noise in the process. This is especially helpful as the number of variants grows. As discussed in Sections 4.1 and 4.2, this supports conclusions like 'edge choice matters far less than node or text choice', which in a practitioner's setting can be as useful as simply having the most performant construction.

    B.: Yes, we have now run full scale evaluations across all five datasets, summarized below and reported per variant in Appendix C. Agreement is slightly weaker than at the subset level: Spearman ρ against full-scale rankings is 0.75–0.90 (mean 0.82) versus the 0.97 proxy mean in Table 3, though all five remain statistically significant. Most of the reordering comes from the edge axis, since attribute-matching edges grow quadratically with graph size: node ordering is preserved on all five datasets and text ordering on four of five, while the edge axis is where the reordering concentrates. Score magnitudes shift, but the ordering of node choices holds throughout.

    C: Yes, thank you for the baseline suggestion; we have added it! Top-1 regret against the full-scale optimum is 0 on three of five datasets (History, ArXiv and Electronics select it outright), with a mean of 0.9 points across all five and top-3 regret of at most 0.1 points (out of 100). We had followed the closest related work, which reports rank correlation rather than oracle regret, but agree this framing is more informative and have added it to Section 4.3.

DatasetFull pool (rows)1k subsamplePaper ρFull-pool ρPaper scoreFull-pool score
History20,7754.8%0.970.8333.2 ± 1.955.5 ± 2.5
Sports50,0002.0%0.980.7529.0 ± 13.761.9 ± 8.9
ArXiv50,0002.0%0.970.7863.1 ± 18.574.4 ± 5.8
Electronics27,4253.6%0.980.9038.1 ± 21.663.2 ± 24.3
Toys29,0643.4%0.950.866.0 ± 4.148.1 ± 3.9
  1. GraphRAG Robustness

    A.: On validating the questions, we generate 300 questions per dataset (233 for History) across three subtypes, and our current checks are the faithfulness gate (Section 3.2.2) and manual review of a small random subset, now written out in Appendix D. We agree an independent evaluator agent over a larger sample is a stronger check; we are running it now and will report it in that same appendix during the discussion period. On separating retrieval from generation, we do save faithfulness, answer relevance and context relevance separately for every question and variant, and point to context relevance in particular: it scores the retrieved passages rather than the generated answer, so it is already a retrieval-only measurement. It is also the axis that separates constructions best, with a between-variant spread over within-variant noise of 0.268 against 0.221 for the composite and 0.113 for faithfulness. We report the composite in the body because it is the lowest variance score of the four (mean within-variant SD of 0.249, against 0.427 for context relevance) and because it is the standard RAGAS summary, but we have added the per-axis rankings to Appendix D so that the retrieval-only view is directly available. It is also worth noting that the graph variant is the only thing we change between runs, so any difference in these scores is attributable to construction.

    B.: To your question of whether the selection holds under alternative evaluations, partly: ranking by context relevance alone picks the same construction as the composite on 2 of 5 datasets, and on the other three the node choice is unchanged while the edge and text choices move. This is consistent with our broader finding that the node axis carries most of the signal and the edge axis is comparatively flat, but we agree it is a limitation of the composite and now report it. Overall the optimal ordering depends partly on the scoring method rather than being universal, and we have added these results to the main body as Section 4.4, with the per-axis tables in Appendix D.

We will upload the revised and reformatted PDF on Monday 11/16.

Thank you for the detailed response and the additional full-scale evaluations. The full-pool correlations and top-1/top-3 regret results substantially strengthen the evidence for practical utility.

I also appreciate the clarification that the decision tree is intended primarily for interpretability rather than improving selection over direct proxy-mean ranking; this positioning should be made explicit in the revised manuscript. However, the GraphRAG analysis shows that context-relevance ranking selects the same construction on only two of five datasets, and the independent-judge validation is not yet available. Thus, my concern about evaluation-dependent graph selection remains only partially resolved.

At this stage, I therefore plan to retain my weak-reject score, while encouraging the authors to include these new analyses and clearly discuss the remaining limitations in the revision.

Comments on the paper

1 comment

We thank all three reviewers for their engagement and follow-up comments. As Reviewers JxMn and 6riB noted, the full-scale evaluation, regret analysis, and cross-architecture results strengthen the paper. Although our earlier comment mentioned uploading a revised PDF on November 16, we would rather integrate the results we described carefully (structure-blind baselines, per-axis GraphRAG breakdowns, and the remaining checks and limitation discussions you highlighted) than upload a partial version now. Beyond specific fixes, this discussion has also been valuable in sharpening how we approach a foundational question that work in TAG learning has not yet treated systematically: how to define and evaluate what an optimal graph construction looks like in practice. We believe our framing takes an early and important step toward answering it, and will incorporate these new, stronger results and a clearer discussion of the limitations, ablations and considerations you raised into our paper. We are grateful for the detailed feedback, which has substantially improved our understanding of how to position our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.