acceptodds
Under review as a conference paper at ICLR 2027

From Missing Relations to Reliable Structure: Cross-View Relation Recovery and Latent Structure Estimation for Incomplete Multi-View Learning

Abstract

In real-world applications, multi-view data are often incomplete because some views are unavailable for some samples. Although existing incomplete multi-view learning methods have made substantial progress in feature recovery and structural modeling, view missingness can also cause relational information to disappear when relation endpoints are unavailable, leaving the underlying structure incomplete. To address this issue, we propose Reliable Latent Structure Estimation (ReLaS), a relation-centric framework that explicitly restores missing relational structure and learns a latent consensus graph from incomplete multi-view observations. For relations that remain directly observable, ReLaS estimates their relation evidence from endpoint availability and cross-view edge confirmations, while characterizing their reliability. Relations rendered unobservable by missing endpoints are explicitly recovered through cross-view structural correspondence, with the reliability of recovered relations further calibrated in a self-supervised manner. Both direct and recovered relation evidence are then integrated under relation-profile consistency to obtain the latent consensus graph. Experiments on multiple benchmark datasets under different view-missing rates demonstrate the effectiveness of ReLaS in restoring relational structure and improving semi-supervised classification performance.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
incomplete multi-view learning, graph structure learning, semi-supervised learning, cross-view relation recovery, reliability estimation

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_s7JAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ReLaS, a framework for incomplete multi-view learning that focuses on recovering missing relations between samples. It includes three main components: direct relation estimation, cross-view relation recovery, and latent consensus graph learning. The method uses information from overlapping views to recover missing relations and estimate their reliability. Experiments on eight benchmark datasets show promising classification results, and ablation studies evaluate the contributions of different components.

Critical strengths

1.The paper addresses an important problem by considering missing relations, not just missing features.

2.The framework combines relation recovery and reliability estimation, and the experiments cover multiple datasets and missing-view rates.

Critical weaknesses

1.The paper does not clearly explain how it differs from existing relation recovery methods.

2.The method relies on overlapping nodes across views, but its limitations under low overlap or inconsistent structures are not well studied.

3.The reliability calibration method may not accurately reflect the correctness of recovered relations.

4.The experiments lack sufficient statistical analysis, such as results across repeated runs and significance tests.

5.The paper provides limited analysis of computational cost and scalability to large datasets.

Questions for authors

Please refer to "Critical Weaknesses".

2 comments
AuthorsOpenReview ↗

Response to Reviewer s7JA (Part 1/2)

Thank you for recognizing the importance of relation-level incompleteness and the combination of recovery and reliability estimation in ReLaS. The comments have helped clarify the methodological distinctions and strengthen the experimental presentation. Relevant revisions are highlighted in red.

1. Distinction from existing relation recovery methods.

Existing approaches address different aspects of incomplete multi-view structure. RTRGC studies similarity-graph completion, ASCR learns and completes anchor-sample similarities, while SFCR and CAL exploit cross-view affinities for structural modeling. More recent methods, including AGF-TI and RCIMVC, further investigate structural information recovery.

ReLaS focuses on the observability and reliability of pairwise evidence under view missingness. DRE distinguishes observation opportunities OeO_e from edge confirmations CeC_e and estimates (qe,πe)(q_e,\pi_e). CRR uses bridge-relative structural correspondence to recover unconfirmed relations and calibrates (re,ηe)(r_e,\eta_e) through controlled pseudo-missingness. LCGL integrates both evidence sources with reliability-weighted fidelity and relation-profile consistency to learn the latent consensus graph. The contribution lies in jointly modeling direct evidence, uncertain recovery, and reliability-aware graph estimation, rather than treating relation recovery as graph completion alone.

Revision: Section 2 (Related Work) explicitly discusses these distinctions; Sections 3.1–3.3 detail the corresponding mechanisms.

2. Low overlap and inconsistent cross-view structures.

CRR establishes structural correspondence through shared bridge samples Buv=Ωu∩Ωv\mathcal B_{uv}=\Omega_u\cap\Omega_v. Bridge-guided candidate generation transfers local neighborhoods, and controlled pseudo-missingness calibrates recovered-relation scores. Only candidates with positive calibrated reliability ηe>0\eta_e>0 are admitted into LCGL, reducing the influence of unsupported transferred evidence.

Importantly, ReLaS does not rely solely on CRR. With limited shared samples, fewer informative cross-view correspondences may be available, but DRE can still estimate the strength and reliability of directly observable relations, and LCGL can learn from the available evidence. If a view pair has no shared bridge samples, recovery through that pair is unavailable; however, DRE and LCGL can still operate on existing direct relations. Thus, the framework retains a direct-evidence pathway when cross-view transfer is limited.

The expanded Fig. 4 reports relation loss and recovery recall on all eight datasets over VMRs of 0.1–0.8. Although recovery rates vary by dataset, CRR continues to recover missing relations as view missingness increases. This analysis complements the method's reliability-aware selection mechanism. The revised conclusion also identifies sufficient bridge information and informative structural correspondence as favorable conditions for ReLaS.

Revision: Section 3.2 describes bridge-guided recovery; Section 4.4 and the updated Fig. 4 extend the relation-level analysis to eight datasets; Section 5 further clarifies the applicability conditions.

3. Reliability calibration and recovered-relation correctness.

Equation (13) does not label unconfirmed relations as semantically incorrect. Its zero target indicates that a specific confirmation-loss event did not occur under controlled pseudo-missingness, not that the underlying relation is false. Isotonic regression calibrates recovered scores against these events; ηe\eta_e measures the relative trustworthiness of recovered evidence beyond the pseudo-recovery base rate, rather than a semantic-correctness probability.

The new Table 3 supports the practical value of reliability-aware recovery. At VMR =0.5=0.5 and LAR =0.1=0.1, Full CRR outperforms No CRR and CRR (Oe=0O_e=0 only) in ACC and Macro-F1 on all eight datasets. Relative to strict-only recovery, the dataset-averaged gains are 0.73 and 0.76 percentage points, respectively. These results demonstrate the downstream benefit of admitting informative, reliability-filtered relations beyond the strictly unobservable set.

Revision: Section 3.2 clarifies pseudo-missingness targets and ηe\eta_e; Section 4.3 and the newly added Table 3 provide the CRR ablation.

AuthorsOpenReview ↗

Response to Reviewer s7JA (Part 2/2)

4. Repeated-run statistics and performance variability.

Thank you for highlighting the need for more complete statistical reporting. In addition to the main ACC/Macro-F1 comparison in Table 1, the revised Appendix B.3 (Tables 4 and 5) presents ACC and Macro-F1 means and standard deviations across all eight benchmark datasets at VMR =0.3=0.3, 0.50.5, and 0.70.7, with 10% labeled samples.

These supplementary results display performance variability alongside average classification results, providing additional information on robustness across experimental settings. The main manuscript also presents results over a wider range of VMRs in Fig. 3 and component-level comparisons in Tables 2 and 3.

Revision: Appendix B.3 (Additional Classification Results), including Tables 4 and 5, has been added.

5. Computational cost and scalability.

ReLaS employs a sparse computational design. CRR generates relation candidates by transferring local neighborhoods through shared bridge nodes rather than exhaustively enumerating all sample pairs. Reliability-based admission limits the recovered set, while LCGL optimizes admitted undirected-edge variables instead of a dense n×nn\times n affinity matrix. These choices help control the cost of relation recovery and structure estimation.

The new Appendix B.2 (Fig. 7) reports runtime comparisons of nine methods across eight datasets at VMR =0.5=0.5 and LAR =0.1=0.1. ReLaS records the shortest runtime on six datasets and the second-shortest on the remaining two. On ALOI-100 (10,800 samples) and Fashion (10,000 samples), mean runtimes are 33.28 s and 40.22 s, respectively.

Datasets with more than 100,000 samples are not currently available within the experimental resources of this study. Nevertheless, sparse candidate generation, edge-variable optimization, and the observed runtime results support the potential of ReLaS for larger-scale incomplete multi-view learning. Further evaluation on larger datasets will be explored in future work.

Revision: Sections 3.2 and 3.3 describe sparse recovery and edge-based optimization; the newly added Appendix B.2 and Fig. 7 provide runtime comparisons.

Thank you again for the constructive comments. The revisions clarify the novelty and applicability of ReLaS and strengthen the evidence concerning reliability-aware recovery, performance variability, and computational efficiency.

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

This paper proposes ReLaS, which explicitly restores missing relational structure and learns a latent consensus graph to address relational information loss caused by incomplete multi-view data. ReLaS estimates directly observable relations and their reliability through DRE, uses CRR to recover missing relations through cross-view structural correspondence, and fuses these two types of relation evidence through LCGL. The learned latent consensus graph is used for downstream semi-supervised classification. Experimental results show that ReLaS performs well under high view-missing rates.

Critical strengths
  1. Meaningful research problem. This paper distinguishes observable but uncertain relations from unobservable relations caused by missing endpoints. It studies incomplete multi-view learning from the relation level, and the motivation is reasonable.
  2. Complete framework. DRE, CRR, and LCGL are used for relation estimation, relation recovery, and structure fusion, respectively. The overall framework is well organized. The appendix also provides statistical analysis of DRE and convexity analysis of LCGL.
  3. Good experimental results. Table 1 shows that ReLaS achieves the best ACC and Macro-F1 on all eight datasets at VMR = 0.5 and 0.7. The ablation results in Table 2 also support the effectiveness of the three modules.
Critical weaknesses
  1. In Equation (50), missing relations are defined as Oe=0O_e=0. CRR only removes relations with Ce>0C_e>0. This means some recovered relations may not be truly missing, but just unconfirmed. The paper does not show how much each type helps the results.
  2. Equation (13) treats unconfirmed relations as negative samples, while some of them may be true relations. It is unclear whether the estimated reliability can reflect the correctness of recovered relations.
  3. The paper does not explain how dud_u and σu\sigma_u in Equation (9) are calculated, how KAK_A is selected, and how EfullE_{\mathrm{full}} is constructed.
Questions for authors
  1. For the recovered relations in ERE_R, how many satisfy Oe=0O_e=0 and how many satisfy Oe>0O_e>0 with Ce=0C_e=0? How would the results in Table 1 change if only relations with Oe=0O_e=0 were kept?
  2. In Table 2, the firt row (without LCGL, DRE, and CRR) already reaches 66.31 on Scene15 and 86.60 on ALOI-100 at VMR=0.5, higher than all eight competitors in Table 1. How exactly is this baseline built, and were the competing methods tuned under the same protocol?
  3. Fig. 4 reports relation loss ratio and recovery recall for only six datasets, and LandUse-21 and Fashion are not among them. Can the authors report results on LandUse-21 and Fashion?
7 comments
AuthorsOpenReview ↗

Response to Reviewer bvaE (Part 1/2)

Thank you for recognizing the motivation, coherent framework, and experimental performance of ReLaS. The comments have helped clarify the types of recovered relations, reliability calibration, and implementation details. The corresponding changes are highlighted in red in the revised manuscript.

1. Types of recovered relations and their contributions.

Equation (50) defines strictly lost reference relations for evaluation, Elost={e∈Efull:Oe=0}\mathcal E_{\mathrm{lost}}=\{e\in\mathcal E_{\mathrm{full}}:O_e=0\}. CRR, however, intentionally considers two types of unconfirmed relations: (i) strictly unobservable relations (Oe=0O_e=0) and (ii) co-observable but unconfirmed relations (Oe>0,Ce=0O_e>0,C_e=0). The absence of direct KNN confirmation does not imply that a relation is incorrect. Cross-view structural correspondence can reveal useful relations not captured by individual view-specific graphs.

The newly added Table 3 reports the numbers and proportions of both types in the admitted set ER\mathcal E_R across all eight datasets at VMR =0.5=0.5 and LAR =0.1=0.1. Type (ii) accounts for 74.53%–90.66% of admitted relations. For example, CUB has 3,114 type-(i) and 9,112 type-(ii) relations; Fashion has 87,026 and 476,702, respectively.

To assess their contributions, Table 3 compares No CRR, CRR (Oe=0O_e=0 only), and Full CRR. Full CRR outperforms strict-only recovery in both ACC and Macro-F1 on all eight datasets, increasing dataset-averaged ACC and Macro-F1 by 0.73 and 0.76 percentage points, respectively. CUB ACC rises from 75.37% to 77.13%, and Fashion ACC from 87.06% to 88.26%. These findings show that reliability-filtered, co-observable but unconfirmed relations contribute complementary structural information beyond strictly lost-relation recovery.

Revision: Section 3.2 now distinguishes the two relation types. Section 4.3 and the new Table 3 present their composition and corresponding ablation results.

2. Pseudo-negative targets and reliability calibration.

Equation (13) does not treat unconfirmed relations as semantically negative samples. It constructs a self-supervised confirmation-loss target through controlled pseudo-missingness. A target of one indicates that a previously confirmed relation loses all confirmations after pseudo-masking; a target of zero indicates only that this event did not occur. Thus, zero does not imply an incorrect or nonexistent relation.

Isotonic regression calibrates recovered scores against these pseudo-recovery outcomes. The resulting weight ηe\eta_e represents the relative trustworthiness of recovered evidence beyond the pseudo-recovery base rate, rather than a probability of semantic correctness. Only relations with ηe>0\eta_e>0 enter the recovered relation set.

The newly added Table 3 also supports the practical value of this reliability-aware recovery design: incorporating CRR improves classification over No CRR, and Full CRR consistently improves upon strict-only recovery across all eight datasets. The mechanism therefore incorporates useful relational evidence without assuming that unconfirmed relations are false.

Revision: Section 3.2 clarifies the meaning of the pseudo-missingness target and ηe\eta_e. Section 4.3 and Table 3 add the relevant CRR ablation evidence.

3. Bridge construction, anchor selection, and relation profiles.

For an ordered view pair u→vu\rightarrow v, the shared bridge Buv=Ωu∩Ωv\mathcal B_{uv}=\Omega_u\cap\Omega_v contains samples observed in both views. For a source node ii, the anchor set Au→v(i)\mathcal A_{u\rightarrow v}(i) selects the KAK_A bridge nodes most strongly connected to ii in view uu. Their target-view KNN neighborhoods are transferred to generate candidate relations.

In Eq. (9), du(i,a)d_u(i,a) is the cosine distance between L2-normalized PCA representations, clipped to [0,2][0,2]. The scale σu\sigma_u is the median KNN distance among observed samples in view uu, with a lower bound of 10−610^{-6}. Relation profiles apply an exponential similarity transformation over all shared bridge nodes, followed by L2 normalization. A common ordering of bridge nodes allows their inner product to measure cross-view structural correspondence.

The complete-view reference relation graph Efull\mathcal E_{\mathrm{full}} is constructed from complete-view reference information for relation-loss and recovery-recall evaluation; it is not provided to CRR during recovery.

Revision: Section 3.2 clarifies bridge and anchor construction and the distance, scale, and relation-profile definitions in Eq. (9).

AuthorsOpenReview ↗

Response to Reviewer bvaE (Part 2/2)

4. Raw-graph baseline and fairness of comparisons.

The first row of Table 2 is not a random or feature-only baseline. It constructs weighted KNN graphs from observed samples in each view, merges their edges, averages weights for edges supported by multiple views, and performs label propagation with the same 10% labeled samples as the other ablation variants. Thus, even without DRE, CRR, or LCGL, the baseline retains informative multi-view relational structure and provides a meaningful starting point for assessing the proposed modules.

Crucially, full ReLaS further improves this strong baseline. At VMR =0.5=0.5, Table 2 shows that Scene15 ACC rises from 66.31% to 69.50% (+3.19 percentage points), and ALOI-100 from 86.60% to 89.34% (+2.74 points). The progressive ablation shows that the proposed components contribute beyond the initial graph construction.

Regarding fairness, publicly available baseline implementations were adapted to shared datasets, view-missing masks, labeled splits, and ACC/Macro-F1 metrics. Following AGF-TI, Deep Matrix Factorization (DMF) was used for methods requiring complete-view inputs. Consistent tuning principles were adopted, with method-specific hyperparameter search spaces rather than simply using default settings.

For example, on ALOI-100 (VMR =0.5=0.5, LAR =0.1=0.1), changing AGF-TI's parameter ρ\rho from 1000 to 100 increased ACC from 11.42% to 81.08%, while reducing MEGNN's hidden dimension from 64 to 32 improved ACC from 23.69% to 47.30%. On CUB (VMR =0.5=0.5, LAR =0.05=0.05), increasing DCP's learning rate from 10−410^{-4} to 3×10−43\times10^{-4} improved ACC from 53.16% to 54.21%.

These examples show that competing methods received active, method-specific hyperparameter tuning. Some achieved substantial improvements, while others showed more modest gains, reflecting different sensitivities to hyperparameter choices.

Revision: Section 4.1 clarifies the shared experimental protocol. Section 4.3 explicitly explains the all-disabled graph baseline and presents the progressive ablation in Table 2.

5. Additional results for LandUse-21 and Fashion in Fig. 4.

Thank you for this suggestion. The revised Fig. 4 now includes LandUse-21 and Fashion, covering all eight benchmark datasets across VMRs from 0.1 to 0.8.

At VMR =0.5=0.5, LandUse-21 and Fashion have relation-loss ratios of 8.62% and 8.19%, respectively, with corresponding recovery recalls of 48.38% and 16.42%. The expanded figure shows that relation loss grows as view missingness increases, while CRR continues to recover lost reference relations. The variation in recovery recall across datasets also reflects differences in their cross-view structural characteristics.

These additional results provide a more comprehensive assessment of relational degradation and recovery, further supporting the motivation for explicitly modeling relation-level incompleteness.

Revision: Section 4.4 and the updated Fig. 4 now present relation-loss and recovery-recall results for all eight datasets.

Thank you again for the constructive comments. The revised manuscript clarifies the two recovered-relation types, pseudo-missingness calibration, relation-profile construction, and the strong graph baseline, while adding targeted CRR ablations and the requested Fig. 4 results.

Thank authors for the clarifications and the additional results. I have two follow-up questions.

  1. In Section 3.2, authors estimate reliability by hiding one observed view for each eligible sample. In practice, missingness may follow a pattern. For example, one view may often be missing while the others are usually available. Would your method still assign higher weights ηe\eta_e to more reliable recovered edges in this setting?
  2. I still have a concern about the baseline in Table 2. Your response clarifies how it is constructed. Could you also explain why it already outperforms the strongest competing method on Scene15 and ALOI-100 at VMR = 0.5? On ALOI-100 in particular, the baseline alone leads by 4.29 percentage points, and the proposed components add another 2.74 points. This suggests that much of the reported gain on this dataset comes from the baseline itself.

I also encourage authors to release the code and experimental configurations for both ReLaS and the baselines alongside the final version of the paper. I know time is limited, so please feel free to keep your responses brief.

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

This paper addresses semi-supervised classification in incomplete multi-view learning. It bypasses traditional feature imputation and directly tackles the relational incompleteness induced by missing views. When a shared view is missing, direct relations between samples become unobservable; existing structural methods often couple structure recovery with feature imputation, failing to fill this structural gap directly and reliably. The proposed workflow features three key modules. First, Direct Relation Estimation uses a Beta posterior model to estimate confidence for observable direct relations. Second, Cross-View Relation Recovery utilizes bridge nodes and controlled pseudo-missingness to generate self-supervised signals, calibrating the reliability of recovered relations via isotonic regression. Third, Latent Consensus Graph Learning integrates both direct and recovered relations into a quadratic objective with Laplacian regularization to construct a latent consensus graph for label propagation. Experimental findings demonstrate that the proposed method achieves superior performance with graceful degradation as the missing rate increases, confirming the complementarity of direct estimation and relation recovery. Ablations confirm the necessity of both modules, while case analyses quantitatively illustrate the severe relation loss under missingness and the method's effectiveness in successfully recovering these lost relations.

Critical strengths

ReLaS attempts to address the direct recovery of cross-view relations that become completely unobservable due to missing views within the original relational space. The method treats relations themselves as the primary objects of modeling, estimation, and recovery, and introduces a reliability mechanism to control the uncertainty inherent in the recovery process.

Critical weaknesses

ReLaS takes "relational incompleteness" as its core problem, but the paper assumes by default that relation-level recovery is necessary and feasible for all scenarios, lacking a critical examination of the problem's boundaries. The solution paths for IMVC cover multiple levels, including the observation, representation, relation, and decision levels. Consequently, the boundary of applicability between this method and the feature imputation route remains ambiguous.

Questions for authors
  1. Why are IMVC methods based on Transformers or contrastive learning not discussed?

  2. In scenarios with extremely low view missing rates or highly strong structural independence between views, do the benefits of relation recovery outweigh the computational overhead introduced by candidate pair enumeration and isotonic regression?

  3. Is the proposed method applicable to large-scale datasets with over 100,000 samples?

2 comments
AuthorsOpenReview ↗

Response to Reviewer YcuP (Part 1/2)

Thank you for the thoughtful and constructive comments. These suggestions have helped improve the clarity, completeness, and presentation of ReLaS. The concerns are addressed individually below, and the corresponding revisions have been incorporated into the revised manuscript, with all changes highlighted in red.

1. Applicability of relation-level recovery and distinction from feature imputation.

Incomplete multi-view learning can be addressed at different levels, including feature reconstruction, representation learning, and structural modeling. These approaches focus on different aspects of missing information and are not mutually exclusive.

As discussed in the revised Section 2 (Related Work), existing feature-recovery methods primarily reconstruct missing observations or learn complete latent representations. In contrast, structure-oriented methods, such as RTRGC and ASCR, investigate similarity graph completion and anchor-based structural learning, while AGF-TI and RCIMVC further explore the recovery of structural information from incomplete observations.

Building upon these developments, ReLaS focuses specifically on the availability and reliability of pairwise relational evidence under view missingness. Its distinction lies in three complementary components:

  • DRE explicitly separates pairwise observation opportunities from positive edge confirmations, providing relation-specific evidence estimates and reliability weights.
  • CRR exploits shared bridge samples and cross-view structural correspondence to recover unconfirmed relations, with reliability further calibrated through controlled pseudo-missingness.
  • LCGL integrates direct and recovered evidence through reliability-aware weighting and relation-profile regularization to estimate the latent consensus graph.

Rather than requiring reconstructed feature vectors, this formulation directly addresses the structural information needed for graph-based semi-supervised classification. Consequently, ReLaS offers a complementary approach to feature imputation, particularly for tasks where sample connectivity and relational structure are central to downstream prediction.

Revision: Section 2 has been expanded to clarify the distinctions between ReLaS and existing feature-recovery and structural-learning methods. Section 5 (Conclusion) further highlights the applicability of ReLaS to graph-based learning tasks.

2. Discussion of Transformer-based and contrastive learning methods.

Transformer-based and contrastive approaches were already included in the original comparative experiments. Specifically, CoGFormer employs a graph Transformer to capture structural dependencies and cross-view consensus, whereas DCP adopts dual contrastive learning to improve the consistency of multi-view representations.

These approaches provide effective mechanisms for learning cross-view dependencies. However, their modeling objectives differ from the relation-centric formulation of ReLaS. Rather than primarily learning consistent representations or structural dependencies, ReLaS explicitly distinguishes observable relation evidence from unconfirmed relations affected by view missingness. It then combines reliability-aware direct relation estimation, cross-view relation recovery, and latent consensus graph estimation within a unified framework.

Furthermore, both CoGFormer and DCP have been evaluated alongside ReLaS in the original comparative experiments. The results in Table 1 demonstrate the competitiveness of ReLaS across eight benchmark datasets under different view-missing rates.

Revision: The revised Section 2 explicitly discusses Transformer-based and contrastive approaches, including CoGFormer and DCP, and further clarifies their methodological differences from ReLaS.

AuthorsOpenReview ↗

Response to Reviewer YcuP (Part 2/2)

3. Benefits of relation recovery under different missing rates and computational efficiency.

The effectiveness of ReLaS comes from the complementary contributions of DRE and CRR, rather than relying exclusively on relation recovery.

At relatively low view-missing rates, more pairwise relations remain directly observable. Under these conditions, DRE can exploit the available structural evidence by distinguishing observation opportunities from actual edge confirmations and estimating their reliability. Thus, reliable direct relation estimation remains an important component even when relatively few relations require recovery.

As view missingness increases, more pairwise relations become unobservable, making CRR increasingly relevant. The updated Fig. 4 presents relation loss ratios and recovery recalls across all eight benchmark datasets under VMRs ranging from 0.1 to 0.8. The results show that relation loss increases substantially with missingness, while CRR continues to recover missing relational evidence. For example, on COIL20, the relation loss ratio increases from 8.53% to 21.49% as VMR rises from 0.5 to 0.8, while recovery recall remains around 90%.

Regarding computational efficiency, ReLaS adopts several sparse and reliability-aware designs. First, CRR employs bridge-guided sparse candidate generation, transferring local neighborhoods through shared bridge samples instead of exhaustively enumerating all possible sample pairs. Second, isotonic regression calibrates scalar recovered-relation scores using self-supervised pseudo-missingness signals. Third, only relations with positive calibrated reliability are admitted into the recovered relation set. Finally, LCGL optimizes nonnegative variables associated with admitted undirected edges rather than a dense sample-by-sample matrix.

These designs help control the computational cost of structural recovery and estimation. The revised Appendix B.2 (Fig. 7) further presents runtime comparisons of nine methods across all eight benchmark datasets. At VMR = 0.5 and LAR = 0.1, ReLaS achieves the shortest recorded runtime on six datasets and the second-shortest on the remaining two, demonstrating its competitive computational efficiency.

Overall, DRE and CRR provide complementary benefits under different missingness conditions, while sparse candidate generation and edge-based optimization support efficient practical computation.

Revision: Section 4.4 presents the updated relation degradation and recovery analysis in Fig. 4. Sections 3.2 and 3.3 describe the sparse recovery and optimization mechanisms, while Appendix B.2 provides the additional runtime comparison in Fig. 7.

4. Scalability to datasets with more than 100,000 samples.

The computational design of ReLaS provides promising potential for extension to larger-scale datasets. Currently, datasets containing more than 100,000 samples are not available within the experimental resources accessible to this study, so experiments at this scale have not yet been conducted.

Specifically, CRR uses bridge-guided sparse candidate generation rather than exhaustive all-pairs enumeration, while LCGL performs optimization over admitted relation variables instead of constructing and optimizing a dense affinity matrix. These design choices reduce the need for dense pairwise computations and provide a favorable foundation for scaling to larger datasets.

The empirical results also support the practical efficiency of this framework. ReLaS has been evaluated on ALOI-100 (10,800 samples) and Fashion (10,000 samples), requiring average runtimes of only 33.28 s and 40.22 s, respectively. The runtime results across all eight benchmark datasets further demonstrate that the proposed structural estimation framework can be implemented efficiently.

Although experiments beyond 100,000 samples have not yet been conducted, the sparse computational design, edge-based optimization formulation, and observed runtime performance collectively support the potential of ReLaS for large-scale incomplete multi-view learning. Further large-scale evaluation and optimization will be explored in future work.

Revision: Sections 3.2 and 3.3 describe the sparse candidate-generation and edge-based optimization strategies. Appendix B.2 reports additional runtime experiments, including results on ALOI-100 and Fashion.

Thank you again for the valuable suggestions. The revisions provide a clearer explanation of the methodological distinctions, complementary roles of direct relation estimation and recovery, and computational efficiency of ReLaS.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.