acceptodds
Under review as a conference paper at ICLR 2027

From Unique to Unified: Attribute-Centric Self-Supervised Representation Learning for Mixed-Type Tabular Data Clustering

Abstract

Mixed-type tabular data are common in scientific and real-world applications, yet learning reliable representations remains challenging because their attributes encode different relational semantics. Current self-supervised paradigms commonly derive supervision from inter-instance similarity, but such relations can be insufficient or unstable for heterogeneous tabular data. Moreover, numerical, nominal, and ordinal attributes encode fundamentally different notions of similarity. We therefore exploit inter-attribute dependencies as a complementary source of self supervision and follow a preserve-before-unify principle, in which type-specific semantics are first retained and subsequently integrated into a shared representation. Based on this principle, we propose Faithful Unification through Semantics-Preserving Encoding (FUSE), an attribute-centric self-supervised framework for faithfully unifying mixed attributes. Its Type-Faithful Attribute Formation (TAF) module aligns perturbation and recovery objectives with the relational properties of each attribute type, preserving their unique semantics before fusion. The Cross-Attribute Relational Unification (CRU) module further captures within-sample cross-attribute dependencies through target-exclusive contextual prediction, providing relational supervision for a unified representation. Comprehensive experiments across diverse mixed-type benchmarks demonstrate the efficiency and superiority of FUSE.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Mixed-type data clustering, Self-supervised learning, Representation learning, Tabular data, Heterogeneous attributes

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XqZ1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FUSE, an attribute-centric self-supervised representation learning framework for mixed-type tabular data clustering. The framework follows a preserve-before-unify principle to accommodate the distinct semantics of numerical, nominal, and ordinal attributes. It consists of two main components: Type-Faithful Attribute Formation (TAF), which preserves type-specific characteristics through tailored perturbation and recovery objectives, and Cross-Attribute Relational Unification (CRU), which captures dependencies among attributes through target-exclusive contextual prediction. Experiments across multiple mixed-type datasets report improvements in clustering performance compared with existing methods.

Critical strengths

The paper addresses an important challenge in clustering mixed-type tabular data by explicitly considering the distinct semantics of numerical, nominal, and ordinal attributes. The preserve-before-unify principle provides a well-motivated foundation for the proposed framework. The combination of type-specific perturbation and recovery objectives with cross-attribute contextual prediction offers a systematic approach to learning unified representations without relying on predefined instance-level similarities. Furthermore, the evaluation across diverse mixed-type datasets suggests the effectiveness and potential applicability of the proposed method.

Critical weaknesses

The novelty of the proposed framework needs further clarification, particularly regarding how the TAF and CRU modules differ from existing type-specific encoding and masked attribute prediction approaches. Although the preserve-before-unify strategy is well motivated, stronger empirical evidence is needed to demonstrate the distinct contributions of its components. Furthermore, the robustness of the learned representations under missing values, noisy attributes, and imbalanced attribute types remains unclear. A more detailed analysis of computational complexity and scalability would also help assess the practical applicability of FUSE to large-scale tabular datasets.

Questions for authors
  1. How do the proposed TAF and CRU modules differ from existing type-specific encoding and masked attribute prediction methods, and what are their individual contributions to clustering performance?

  2. How robust is FUSE when handling missing values, noisy attributes, or datasets with highly imbalanced proportions of numerical, nominal, and ordinal attributes?

  3. How does the computational complexity of FUSE compare with existing mixed-type clustering methods, and how well does it scale to datasets with large numbers of samples and attributes?

2 comments
AuthorsOpenReview ↗

Additional Evidence on FUSE's Design and Robustness, with Scalability Analysis

W1/Q1 [Distinctions and Contributions of the TAF and CRU Modules]: TAF avoids homogenizing heterogeneous attributes into a unified categorical/numerical reconstruction task, instead designing type-faithful learning signals to capture numerical continuity, nominal category relationships, and ordinal ordering. Building on this, the CRU module dynamically rotates target attributes and predicts each from the remaining attributes, encouraging unified representations to capture sample-wise attribute relationships, where identical attribute values may convey different relational information across contexts. These complementary mechanisms provide richer information for mixed-type clustering.

The individual contributions of TAF and CRU are demonstrated by the ablation results (main paper, Section 4.3, Table 2). We additionally conduct comparisons with targeted design variants (see Table 1 below). Specifically, replacing TAF's type-specific corruption-recovery with uniform masking-based recovery (Generic Recovery) degrades clustering performance, supporting the effectiveness of type-matched supervision. Applying both objectives to a shared representation (Shared-space Objectives) also reduces performance, supporting the preserve-before-unify design.

DatasetMetricFUSEGeneric RecoveryShared-space ObjectivesACC.7806±.0167.7659±.0165.7452±.0512HFNMI.2297±.0326.1998±.0395.1865±.0477ARI.3089±.0356.2765±.0375.2437±.0954ACC.6639±.0194.4519±.0333.4190±.0213ANLNMI.3902±.0210.1621±.0123.1744±.0063ARI.2347±.0229.0363±.0111.0243±.0060ACC.9767±.0047.8659±.1212.7324±.0577DGNMI.9496±.0076.8409±.0794.6795±.0434ARI.9580±.0067.8179±.1349.6564±.0814ACC.7593±.0123.7480±.0062.7502±.0027BCNMI.1415±.0079.1163±.0075.1157±.0029ARI.2451±.0177.2174±.0104.2186±.0025\begin{array}{l|c|ccc}\hline\text{Dataset}&\text{Metric}&\text{FUSE}&\text{Generic Recovery}&\text{Shared-space Objectives}\\\\\hline&\text{ACC}&\mathbf{.7806}\pm.0167&.7659\pm.0165&.7452\pm.0512\\\\\text{HF}&\text{NMI}&\mathbf{.2297}\pm.0326&.1998\pm.0395&.1865\pm.0477\\\\&\text{ARI}&\mathbf{.3089}\pm.0356&.2765\pm.0375&.2437\pm.0954\\\\\hline&\text{ACC}&\mathbf{.6639}\pm.0194&.4519\pm.0333&.4190\pm.0213\\\\\text{ANL}&\text{NMI}&\mathbf{.3902}\pm.0210&.1621\pm.0123&.1744\pm.0063\\\\&\text{ARI}&\mathbf{.2347}\pm.0229&.0363\pm.0111&.0243\pm.0060\\\\\hline&\text{ACC}&\mathbf{.9767}\pm.0047&.8659\pm.1212&.7324\pm.0577\\\\\text{DG}&\text{NMI}&\mathbf{.9496}\pm.0076&.8409\pm.0794&.6795\pm.0434\\\\&\text{ARI}&\mathbf{.9580}\pm.0067&.8179\pm.1349&.6564\pm.0814\\\\\hline&\text{ACC}&\mathbf{.7593}\pm.0123&.7480\pm.0062&.7502\pm.0027\\\\\text{BC}&\text{NMI}&\mathbf{.1415}\pm.0079&.1163\pm.0075&.1157\pm.0029\\\\&\text{ARI}&\mathbf{.2451}\pm.0177&.2174\pm.0104&.2186\pm.0025\\\\\hline\end{array} Table 1. Ablations on objective design and representation-level organization.

W2/Q2 [Robustness Evaluation]: FUSE addresses imperfect inputs through type-specific perturbations and cross-attribute context. Specifically, TAF employs type-specific perturbations to simulate missing values, measurement errors, and ordinal rating errors. The resulting recovery tasks provide self-supervised learning signals while encouraging structure preservation under corruption. CRU incorporates complementary contextual information, reducing reliance on individual attributes. Consistently, FUSE shows robustness to attribute noise (main paper, Fig. 4(a)). We additionally evaluate missingness and attribute-type imbalance:

1) Missing values: We compare FUSE with SubTab, T-JEPA, and MGBCC under 5%–40% random missingness. All methods are trained on clean data and evaluated on incomplete inputs. FUSE performs best across all tested rates on CMC and DG and remains competitive on BC. Full results and robustness curves are provided in Appendix C.7 of the revised paper.

2) Imbalanced Attribute Types: FUSE's type-wise averaged recovery losses limit the dominance of majority attribute types. We further conduct a controlled experiment on synthetic data (2,000 samples, 30 attributes), varying numerical/nominal/ordinal proportions while keeping clustering information fixed. FUSE maintains 84.01%–87.15% ARI across imbalanced settings, comparable to 85.33% under balanced proportions. Under nominal dominance (1/28/1), SubTab and MGBCC drop from 44.03% to 12.03% and 70.09% to 36.38%, respectively. These findings, together with FUSE's first-place performance on ordinal-dominant DG and nominal-dominant ANL (main paper, Table 1), support its effectiveness under attribute-type imbalance.

W3/Q3 [Computational Complexity and Scalability]: Our FUSE supports scalable training without constructing pairwise relationships across all samples. CRU controls the computational cost of contextual prediction by limiting the number of target attributes selected per batch. Consequently, training complexity scales linearly with the number of samples under fixed model and training settings (main paper, Section 3). This is further supported by the near-linear runtime growth observed over 1k–60k samples (empirical exponent 0.991; main paper, Fig. 4(b)). Runtime also increases smoothly across 10–120 attributes. We have expanded the discussion of the scaling conditions and computational trade-offs in the revised paper (Section 3.4 and Appendix A).

AuthorsOpenReview ↗

Follow-up on Our Responses

Dear Reviewer XqZ1,

We sincerely appreciate your positive assessment of our work.

In our rebuttal, we responded by clarifying our FUSE’s novelty and the contributions of its components (W1/Q1), extending our robustness evaluations (W2/Q2), and elaborating on computational complexity and scalability (W3/Q3).

If you have any questions or would like to discuss any of these points, we would be pleased to engage further.

Best regards,

The Authors

Reviewer_s3oHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a self-supervised representation learning framework, FUSE, for clustering mixed-type tabular data. The framework consists of Type-Faithful Attribute Formation (TAF), which applies type-specific perturbation and recovery objectives to numerical, nominal, and ordinal attributes, and Cross-Attribute Relational Unification (CRU), which learns unified representations through target-exclusive contextual prediction. Experiments on eleven tabular benchmarks show strong clustering performance against traditional clustering and self-supervised representation learning methods.

Critical strengths
  • Addresses a practical challenge in mixed-type tabular representation learning, where numerical, nominal, and ordinal attributes exhibit distinct relational semantics.

  • Proposes a coherent framework that separates type-specific representation formation from cross-attribute dependency modeling, with learning objectives tailored to different attribute types.

  • Demonstrates strong clustering performance across diverse datasets, supported by component ablations, robustness experiments, and scalability analysis.

Critical weaknesses
  • The problem motivation and methodological novelty are not sufficiently established in the broader context of tabular representation learning. Existing methods such as TabTransformer and FT-Transformer as well as frameworks like PyTorch Frame, already support type-aware feature encoding and cross-feature interaction. Related table embedding and pretraining methods, such as TURL, TabularBERT, and NAVI, also explore contextual representations and heterogeneous attribute semantics, although they target more generalizable tabular representation settings. The authors are encouraged to expand the related work discussion and clarify what limitations of these approaches motivate the proposed preserve-before-unify design, beyond existing type-aware representation learning techniques.

  • The experimental evaluation primarily relies on downstream clustering performance to support the broader claims of semantic preservation and relational unification. While the clustering results are strong, they do not directly establish whether the learned representations preserve type-specific semantics or capture meaningful cross-attribute dependencies. If possible, though not necessary, additional representation-level analyses and evaluations on other downstream tabular tasks, such as classification or regression, would help substantiate these claims and clarify the model's applicability and extensibility.

  • The necessity of several architectural and optimization choices remains insufficiently justified. Although the ablation studies demonstrate the contributions of TAF and CRU, they do not fully explain why separating their optimization through gradient detachment is preferable to joint training, or whether the proposed design offers clear advantages over simpler type-aware encoding and fusion strategies. Controlled comparisons with joint optimization, conventional type-aware encoders, and early-fusion architectures under comparable self-supervised settings would help isolate the specific benefits of the proposed design.

Questions for authors

Please see weaknesses

3 comments
AuthorsOpenReview ↗

Distinctions from Prior Work, Representation Quality Analysis, and Downstream Generalization

W1 [Related Work and Distinctions Discussion]: Compared with the methods noted by the reviewer, FUSE adopts a preserve-before-unify design that organizes type-specific preservation and cross-attribute learning at different representation levels for mixed-type clustering. We conduct additional ablations and clustering comparisons to support our design, with FUSE achieving higher mean ARI than TabularBERT (ICML'26) on 10/11 datasets.

1) Type-aware Tabular Modeling: TabTransformer and FT-Transformer handle heterogeneous attributes through embedding or tokenization and self-attention, primarily for supervised prediction. In contrast, FUSE incorporates numerical, nominal, and ordinal relational structures into type-matched self-supervised recovery objectives for clustering.

2) Contextual and Heterogeneous Tabular Pretraining: TURL models relational-table context, while NAVI learns attribute semantics through header–value pairs. TabularBERT incorporates numerical structure into self-supervised pretraining. FUSE places type-specific recovery and cross-attribute prediction at pre-fusion and unified representation levels, respectively.

The additional ablations (see our separate response to W3) and comparison with TabularBERT (revised paper, Appendix C.2) support our preserve-before-unify design, highlighting the value of bridging type-specific attribute semantics and sample-level relational information for mixed-type clustering. We have expanded the related work discussion (revised paper, Section 2) and included the additional results.

W2 [Representation Quality and Downstream Generalization]: FUSE’s learned representations exhibit semantic preservation and relational unification, which we analyze below. We also evaluate their applicability to classification and regression tasks.

1) Type-specific Structural Preservation: To assess whether TAF preserves type-specific structures, we measure Spearman correlations between input-space and representation-space distances. Compared with Generic Recovery, FUSE achieves higher correlations for numerical (0.9845 vs. 0.9463), nominal (0.8727 vs. 0.6644), and ordinal attributes (0.8649 vs. 0.7724), supporting the effectiveness of type-matched recovery.

2) Cross-attribute Relationship Unification: We assess cross-attribute dependency learning by predicting selected attributes from the remaining context. On ANL nominal attributes, FUSE achieves lower prediction CE (0.1978) than w/o CRU (0.2232) and the marginal-frequency baseline (0.7210), supporting CRU's contribution to contextual prediction.

3) Applicability to Other Downstream Tasks: FUSE's representations can support other tabular tasks, as they capture the underlying structure of heterogeneous data through self-supervised objectives. Accordingly, we evaluate FUSE on classification and regression tasks against tabular SSL methods and a direct-input baseline, using a logistic classifier and a ridge regressor. As shown in Table 1 of this response, FUSE achieves the best performance on AC, DRUG, and CA, while remaining competitive on AMES. These results support the generalizability of FUSE’s representations. We have included dataset details (revised paper, Appendix C.1) and additional results (revised paper, Appendix C.8) in the appendix.

\hline \text{Dataset}&\text{Metric}&\text{FUSE}&\text{LogReg/Ridge}&\text{SubTab}&\text{T-JEPA}&\text{ICE-T}\\\\ \hline \text{AC}&\text{AUROC}\uparrow&\mathbf{0.8942}&\underline{0.8719}&0.8659&0.8709&0.8690\\\\ \scriptstyle(690/6/8/0)&\text{Macro-F1}\uparrow&\mathbf{0.8335}&0.8149&0.8076&\underline{0.8208}&0.8134\\\\ \hline \text{DRUG}&\text{AUROC}\uparrow&\mathbf{0.8390}&0.7760&0.7813&0.7739&\underline{0.8107}\\\\ \scriptstyle(1{,}885/12/18/0)&\text{Macro-F1}\uparrow&\mathbf{0.2082}&0.1846&0.1702&\underline{0.1884}&0.1822\\\\ \hline \text{CA}&\text{MAE}\downarrow&\mathbf{41{,}967.83}&49{,}365.06&45{,}748.05&\underline{43{,}558.22}&46{,}248.93\\\\ \scriptstyle(20{,}640/8/1/0)&R^2\uparrow&\mathbf{0.7326}&0.6366&0.6860&\underline{0.7036}&0.6622\\\\ \hline \text{AMES}&\text{MAE}\downarrow&19{,}314.65&\underline{18{,}586.17}&19{,}580.63&19{,}698.13&\mathbf{17{,}426.76}\\\\ \scriptstyle(2{,}930/33/25/21)&R^2\uparrow&\underline{0.8639}&0.8092&0.8218&0.8115&\mathbf{0.8766}\\\\ \hline \end{array}$$ Table 1. Classification (AC, DRUG) and regression (CA, AMES) performance. Dataset statistics: (samples/num./nom./ord.).
AuthorsOpenReview ↗

Ablations on FUSE’s Optimization, Encoding, and Fusion Strategies

W3 [Architectural and Optimization Ablations]: Gradient separation protects the formation of type-specific representations from potential interference, while the preserve-before-unify design enables learned fusion to build upon these representations. We examine 1) optimization and objective placement and 2) encoding and fusion design through controlled comparisons.

Specifically, we remove gradient detachment to allow joint training (w/o Detachment) and move both objectives after fusion (Shared-space Objectives) to assess the effect of objective placement. We then use conventional MLP encoders with one-hot or embedding inputs (w/ One-hot and w/ Embedding), feed preprocessed attributes directly into the fusion module without type-specific encoders (Early Fusion), and evaluate clustering on concatenated type-specific representations (Direct Concat).

As shown in Table 2 below (revised paper, Appendix C.3.3), removing gradient detachment reduces clustering performance, particularly on HF and ANL. Shared-space Objectives leads to substantial performance drops on ANL and DG. These results support the benefits of gradient separation and the preserve-before-unify design.

FUSE outperforms both conventional encoding variants on all four datasets, demonstrating the effectiveness of its type-specific encoding design. Early Fusion reduces performance on HF, ANL, and BC, showing the value of type-specific representation formation before fusion. Direct Concat produces larger drops on HF and DG, demonstrating that learned fusion further improves the clustering utility of these type-specific representations. \begin{array}{c|c|cc}\begin{array}{l|c|c}\hline\text{Dataset}&\text{Metric}&\text{FUSE}\\\\\hline&\text{ACC}&\mathbf{.7806}\pm.0167\\\\\text{HF}&\text{NMI}&\mathbf{.2297}\pm.0326\\\\&\text{ARI}&\mathbf{.3089}\pm.0356\\\\\hline&\text{ACC}&\mathbf{.6639}\pm.0194\\\\\text{ANL}&\text{NMI}&\mathbf{.3902}\pm.0210\\\\&\text{ARI}&\mathbf{.2347}\pm.0229\\\\\hline&\text{ACC}&\mathbf{.9767}\pm.0047\\\\\text{DG}&\text{NMI}&\mathbf{.9496}\pm.0076\\\\&\text{ARI}&\mathbf{.9580}\pm.0067\\\\\hline&\text{ACC}&\mathbf{.7593}\pm.0123\\\\\text{BC}&\text{NMI}&\mathbf{.1415}\pm.0079\\\\&\text{ARI}&\mathbf{.2451}\pm.0177\\\\\hline\end{array}\kern-.5em&\kern-.5em\overset{\text{Objective \\& Gradient Design}}{\begin{array}{cc}\hline\text{Shared-space Objectives}&\text{w/o Detachment}\\\\\hline.7452\pm.0512&.7258\pm.0574\\\\.1865\pm.0477&.1816\pm.0459\\\\.2437\pm.0954&.2125\pm.0893\\\\\hline.4190\pm.0213&.6010\pm.0888\\\\.1744\pm.0063&.3735\pm.0322\\\\.0243\pm.0060&.1999\pm.0564\\\\\hline.7324\pm.0577&.9313\pm.0816\\\\.6795\pm.0434&.9255\pm.0375\\\\.6564\pm.0814&.9050\pm.0921\\\\\hline.7502\pm.0027&.7545\pm.0088\\\\.1157\pm.0029&.1278\pm.0121\\\\.2186\pm.0025&.2312\pm.0164\\\\\hline\end{array}}\kern-.5em&\kern-.5em\overset{\text{Encoding}}{\begin{array}{cc}\hline\text{w/ One-hot}&\text{w/ Embedding}\\\\\hline.7472\pm.0292&.7512\pm.0130\\\\.1744\pm.0338&.1629\pm.0253\\\\.2396\pm.0524&.2428\pm.0259\\\\\hline.5088\pm.0783&.5509\pm.0691\\\\.3243\pm.0282&.3280\pm.0374\\\\.1282\pm.0361&.1512\pm.0328\\\\\hline.8570\pm.0425&.8391\pm.0445\\\\.8609\pm.0184&.8556\pm.0166\\\\.8288\pm.0515&.8227\pm.0427\\\\\hline.7495\pm.0049&.7379\pm.0043\\\\.1187\pm.0103&.1049\pm.0103\\\\.2207\pm.0122&.1993\pm.0108\\\\\hline\end{array}}&\kern{-1em}\overset{\text{Fusion}}{\begin{array}{cc}\hline\text{Early Fusion}&\text{Direct Concat}\\\\\hline.7380\pm.0140&.5452\pm.0142\\\\.1682\pm.0167&.0024\pm.0014\\\\.2231\pm.0254&.0029\pm.0058\\\\\hline.5886\pm.0525&.6132\pm.0324\\\\.3075\pm.0707&.3308\pm.0216\\\\.1518\pm.0567&.1734\pm.0332\\\\\hline.9618\pm.0026&.7412\pm.0013\\\\.9369\pm.0020&.8206\pm.0044\\\\.9370\pm.0037&.6835\pm.0039\\\\\hline.7304\pm.0061&.7196\pm.0283\\\\.0912\pm.0040&.1063\pm.0181\\\\.1818\pm.0067&.1801\pm.0422\\\\\hline\end{array}}\end{array} Table 2. Effects of FUSE's architectural and optimization choices.

AuthorsOpenReview ↗

Follow-up on Our Responses

Dear Reviewer s3oH,

Thank you for your time and thoughtful comments.

In our rebuttal, we have carefully addressed your concerns by clarifying our FUSE's novelty (W1), providing representation-level analyses and evaluations on other downstream tasks (W2), and conducting further ablations on key design choices (W3).

As the discussion period enters its final day, we would be grateful if you could review our responses and let us know whether they have helped address your concerns. We would be willing to discuss any remaining concerns or provide further clarification and supporting evidence.

Best regards,

The Authors

Reviewer_ZC2wOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes the FUSE self-supervised representation learning framework, specifically designing the TAF module to address the instability of traditional sample-level contrastive learning, while designing the CRU for random erasure of attributes and reconstruction of attributes to achieve attribute uniformity. It achieves a significant improvement under multiple baseline comparisons.

Critical strengths

S1: This paper addresses the main challenges of traditional mixed-type table data, through attribute construction and fusion with TAF and CRU respectively. S2: It achieves certain improvement by demonstrating the FUSE framework on multiple datasets.

Critical weaknesses

W1: The article's ablation experiment granularity is too large, Table 2 only shows the results of FUSE framework ablation of TAF and CRU. The data corruption methods for TAF against different types of data should be conducted in a fine-grained manner. At the same time, it is recommended to compare the random attribute erasure design in CRU with the fixed attribute ablation. W2:The claims in the article do not match the actual experimental section. Section 4.1 claims experiments were conducted on 11 benchmarks, but Table 1 and 2 only show 8 datasets. Additionally, the appendix mentioned in the article could not be retrieved. These discrepancies should be carefully reviewed and corrected, and a supplementary explanation of the content of the appendix claimed in the article should be provided.

Questions for authors

See Weaknesses.

2 comments
AuthorsOpenReview ↗

Fine-grained ablations and appendix clarification

W1 [Fine-grained Ablations]

To show TAF's type-specific encoding and corruption–recovery design, we ablate 1) the overall formation design and 2) specific corruption strategies. Specifically, type-specific corruption–recovery is first replaced with a shared scheme, retaining the type-specific encoders (w/ Shared Rec.). These encoders are then replaced with a single MLP to examine whether type-specific encoding is necessary (w/ Shared Enc. & Rec.). The effect of corruption strategy is further examined through Mask-only, Noise-only, and random replacement of visible values (Generic Replacement).

As shown in Table 1 Formation Design\text{Formation Design}, common recovery reduces clustering performance, particularly on nominal-heavy ANL, suggesting that type-adapted corruption–recovery may better preserve attribute semantics. Further encoder sharing lowers performance on ordinal-rich DG and mixed-type AC and DRUG, indicating the benefit of dedicated encoders for handling heterogeneous attribute structures. For BC, both attribute types are discrete, allowing a shared encoder to capture their common characteristics with little additional loss from encoder sharing.

For Corruption Strategies\text{Corruption Strategies}, FUSE outperforms Noise-only, while Mask-only remains competitive, as masking hides target values without leaving residual clues. On AC, Generic Replacement also performs well, possibly due to harder recovery tasks, while FUSE uses local perturbations to balance recovery difficulty and context preservation.

To evaluate the rationale for CRU’s dynamic target selection, target attributes are sampled once and kept fixed throughout training (w/ Fixed Target). Table 1 CRU\text{CRU} reports higher scores for FUSE than fixed targets on all five datasets, suggesting that greater diversity in prediction tasks can help the unified representation capture broader cross-attribute relationships.

\begin{array}{l|c|c} \hline \text{Dataset}&\text{Metric}&\text{FUSE}\\\\ \hline &\text{ACC}&\mathbf{.6639}\pm.02\\\\ \text{ANL}&\text{NMI}&\mathbf{.3902}\pm.02\\\\ &\text{ARI}&\mathbf{.2347}\pm.02\\\\ \hline &\text{ACC}&.8501\pm.01\\\\ \text{AC}&\text{NMI}&.3912\pm.02\\\\ &\text{ARI}&.4898\pm.03\\\\ \hline &\text{ACC}&\mathbf{.9767}\pm.00\\\\ \text{DG}&\text{NMI}&\mathbf{.9496}\pm.01\\\\ &\text{ARI}&\mathbf{.9580}\pm.01\\\\ \hline &\text{ACC}&\mathbf{.7593}\pm.01\\\\ \text{BC}&\text{NMI}&\mathbf{.1415}\pm.01\\\\ &\text{ARI}&\mathbf{.2451}\pm.02\\\\ \hline &\text{ACC}&\mathbf{.4417}\pm.01\\\\ \text{DRUG}&\text{NMI}&.1020\pm.01\\\\ &\text{ARI}&\mathbf{.1325}\pm.01\\\\ \hline \end{array} & \overset{\text{TAF Formation Design}}{ \begin{array}{cc} \hline \text{w/ Shared Rec.}&\text{w/ Shared Enc. \\&Rec.}\\\\ \hline .4519\pm.03&.4276\pm.04\\\\ .1621\pm.01&.1560\pm.01\\\\ .0363\pm.01&.0234\pm.01\\\\ \hline .8209\pm.01&.8093\pm.01\\\\ .3263\pm.03&.2989\pm.02\\\\ .4113\pm.03&.3815\pm.02\\\\ \hline .8659\pm.12&.7547\pm.11\\\\ .8409\pm.08&.7461\pm.08\\\\ .8179\pm.13&.7118\pm.13\\\\ \hline .7480\pm.01&.7466\pm.01\\\\ .1163\pm.01&.1198\pm.01\\\\ .2174\pm.01&.2181\pm.02\\\\ \hline .4133\pm.02&.3901\pm.02\\\\ .1075\pm.01&\mathbf{.1189}\pm.00\\\\ .1140\pm.02&.0932\pm.01\\\\ \hline \end{array} } & \overset{\text{TAF Corruption Strategies}}{ \begin{array}{ccc} \hline \text{Mask-only}&\text{Noise-only}&\text{Generic Replacement}\\\\ \hline .6469\pm.07&.5900\pm.07&.6030\pm.10\\\\ .3829\pm.02&.3366\pm.04&.3765\pm.03\\\\ .2298\pm.05&.1765\pm.05&.2068\pm.06\\\\ \hline \mathbf{.8554}\pm.01&.8316\pm.01&\mathbf{.8554}\pm.01\\\\ .4014\pm.02&.3508\pm.03&\mathbf{.4033}\pm.01\\\\ \mathbf{.5044}\pm.02&.4393\pm.03&\mathbf{.5044}\pm.02\\\\ \hline .9246\pm.09&.8268\pm.12&.9229\pm.09\\\\ .9148\pm.06&.8441\pm.11&.9122\pm.05\\\\ .8958\pm.10&.8027\pm.15&.8937\pm.10\\\\ \hline .7516\pm.01&.7357\pm.01&.7538\pm.01\\\\ .1228\pm.01&.0964\pm.01&.1248\pm.01\\\\ .2251\pm.02&.1909\pm.02&.2290\pm.01\\\\ \hline .4406\pm.01&.4324\pm.01&.4410\pm.01\\\\ .1073\pm.01&.1027\pm.02&.1016\pm.01\\\\ .1239\pm.01&.1220\pm.00&.1275\pm.00\\\\ \hline \end{array} } & \overset{\text{CRU}}{ \begin{array}{c} \hline \text{w/ Fixed Target}\\\\ \hline .6168\pm.00\\\\ .3174\pm.04\\\\ .1862\pm.03\\\\ \hline .8394\pm.01\\\\ .3627\pm.03\\\\ .4607\pm.04\\\\ \hline .9257\pm.08\\\\ .9214\pm.03\\\\ .8970\pm.08\\\\ \hline .7545\pm.01\\\\ .1248\pm.02\\\\ .2302\pm.03\\\\ \hline .3901\pm.04\\\\ .0981\pm.00\\\\ .0839\pm.02\\\\ \hline \end{array} } \end{array}$$ Table 1. Fine-grained ablation results of TAF and CRU. **W2 [Appendix Clarification]:** The complete results for 11 benchmarks and component ablations are provided in Appendix C.2 (Table 4) and C.3 (Table 5), respectively. We will carefully specify the dataset coverage in the main text and improve the clarity of appendix references.
AuthorsOpenReview ↗

Follow-up on Our Responses

Dear Reviewer ZC2w,

Thank you for your time and thoughtful comments.

In our rebuttal, we have carefully addressed your concerns by conducting fine-grained ablations (W1) and improving the clarity of the revised paper and appendix (W2).

As the discussion period enters its final day, we would be grateful if you could review our responses and let us know whether they have helped address your concerns. We would be willing to discuss any remaining concerns or provide further clarification and supporting evidence.

Best regards,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Genuinely Looking Forward to Reviewers' Feedback

Dear Area Chair and Reviewers,

We thank the Area Chair for coordinating the discussion.

In our rebuttal, we have clarified our approach and conducted additional experiments to further demonstrate our FUSE's effectiveness. We believe our responses sufficiently address the reviewers' concerns.

As the discussion deadline approaches, we would be grateful if the reviewers could review our responses and let us know whether their concerns have been addressed or which issues remain unresolved.

Thank you for your time and consideration.

The Authors

AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair and Reviewers,

As the discussion phase draws to a close, we would like to provide an executive summary of our rebuttal.

Core Motivation and Innovation: In mixed-type data clustering, discrepancies in similarity structures across attribute types, along with their inherent ambiguity, make sample relationships difficult to characterize with predefined metrics. To address this challenge, our FUSE adopts a preserve-before-unify principle for clustering-oriented representation learning. Specifically, the TAF formation module introduces type-matched recovery objectives under tailored corruption strategies, ensuring that self-supervised learning respects the intrinsic structures. Building on this foundation, CRU's contextual prediction uncover cross-attribute dependencies, providing an alternative source of relational self-supervision for clustering. This complementary design grounds cross-attribute relational learning in type-faithful representations, yielding unified representations that reveal underlying sample relationships across heterogeneous attributes.

Supporting Evidence: To address the reviewers' concerns, we conducted additional analyses and evaluations in three aspects: (1) Design Rationale. Fine-grained ablations and controlled comparisons validate component contributions and support our FUSE's design choices (Reviewer ZC2w, W1; Reviewer s3oH, W3; Reviewer XqZ1, W1/Q1). (2) Comparative Evidence. We enriched the related work discussion and compared FUSE with reviewer-mentioned, closely related TabularBERT (ICML’26) to highlight FUSE's performance advantages (Reviewer s3oH, W1; Reviewer XqZ1, W1/Q1). (3) Representation Quality and Applicability. Additional representation analyses examine type-specific preservation and cross-attribute unification, while downstream and robustness evaluations assess the broader applicability of FUSE (Reviewer s3oH, W2; Reviewer XqZ1, W2/Q2).

Manuscript Updates: We have incorporated the above analyses and results into the revised manuscript and appendix, particularly the comparative and ablation studies (Sections 4.2–4.3; Appendices C.2–C.3) and the representation analyses, downstream evaluations, and robustness tests (Appendices C.6–C.8). We also elaborated on the existing complexity analysis and scaling conditions (Appendix A; Reviewer XqZ1, W3/Q3) and made the correspondence between the manuscript and appendix clearer (Sections 4.2–4.3; Reviewer ZC2w, W2).

We believe our rebuttal has addressed the concerns raised and provided more comprehensive supporting evidence. We sincerely thank the Area Chair and reviewers once again for your time, expertise, and constructive comments.

Best regards,

Authors of Paper ID 58560

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.