acceptodds
Under review as a conference paper at ICLR 2027

GRADE: Evidence-Aligned Adaptive Granularity for Structure-Aware Residual Diffusion in Incomplete Multi-View Clustering

Abstract

Incomplete multi-view clustering aims to discover a consistent clustering structure from partially observed multi-view data. However, recovering missing representations still faces two key challenges: (i) relying solely on shared semantics from the observed views can preserve cross-view consistency but cannot fully characterize the local structure specific to the target view; and (ii) view heterogeneity makes it difficult to establish reliable semantic correspondences between local structures across views, hindering the acquisition of query-relevant local priors from the target view. To address these challenges, we propose GRADE. First, we construct local regions for each view at adaptive granularities and use jointly observed members as evidence to establish soft correspondences across views. This extends sparse pairwise evidence to adaptive local regions and enables the retrieval of query-relevant local priors from the target view. Second, we map the shared semantics of the observed views into a target-view base representation. Rather than directly generating the entire missing representation, GRADE uses the retrieved local priors to guide residual diffusion, generating only the target-view-specific local information not captured by the base representation. Experiments on eight multi-view benchmark datasets across multiple missing rates demonstrate the superior overall clustering performance of GRADE.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete Multi-View Clustering, Diffusion Model, Adaptive Granularity

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_efTCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GRADE combines adaptive granularity alignment and structure-aware residual diffusion for incomplete multi-view clustering. Local correspondences and retrieved target-view prototypes guide completion. Experiments cover eight datasets and five missingness settings.

Critical strengths
  1. The paper is well written and easy to follow.
  2. Experimental settings, baseline configurations, and source-code links are clearly provided.
  3. Comparisons, ablations, and efficiency analyses support the method’s effectiveness.
Critical weaknesses
  1. Performance under high missingness needs further explanation. Since alignment depends on co-observed samples, the paper should clarify how reduced evidence affects correspondence, retrieval, and completion quality.
  2. Key design contributions are not fully isolated. Ablations do not directly compare adaptive versus fixed granularity or retrieved versus simpler conditioning. Key ablations and parameter analyses also use only a missing ratio of 0.9 and seed 42, limiting evidence of stability.
Questions for authors

Under higher missingness, is degradation mainly caused by weaker correspondence evidence or fewer training samples for completion? What existing observations support this explanation?

6 comments

We sincerely thank you for your careful review of our work. Your questions and suggestions demonstrate a deep engagement with our method and have helped us further improve the manuscript. We address each comment point by point below and describe the corresponding additional experiments and revisions.

W1:Performance under high missingness needs further explanation. Since alignment depends on co-observed samples, the paper should clarify how reduced evidence affects correspondence, retrieval, and completion quality.

We thank the reviewer for raising this concern. GRADE establishes region correspondences based on co-observed samples and uses them to learn cross-view alignment in the shared space. It then constructs diffusion conditions by matching the query with target-view prototypes. Therefore, as the number of co-observed samples decreases, the statistical evidence for region correspondence becomes limited. The resulting uncertainty may propagate to shared-space alignment, affecting the semantic matching between queries and retrieved prototypes and weakening the structural guidance provided to the diffusion model.

Meanwhile, fewer samples jointly observed in the target view and other views also limit the training information available to the diffusion model. Thus, performance degradation under high missing rates may involve both the reliability of correspondences and retrieval conditions and the sufficiency of diffusion training samples. To further distinguish their effects, we report controlled experiments in Q1 that separately reduce region correspondence evidence, diffusion training samples, and both simultaneously, and analyze their respective impacts on performance.

W2:Key design contributions are not fully isolated. Ablations do not directly compare adaptive versus fixed granularity or retrieved versus simpler conditioning. Key ablations and parameter analyses also use only a missing ratio of 0.9 and seed 42, limiting evidence of stability.

We thank the reviewer for the suggestion. To further distinguish the contributions of key components, we added two ablation variants: replacing granular-ball prototypes with K-means centroids and removing the retrieved structural condition C from the diffusion model. As shown in the table, at a missing rate of 0.9 with seed 42, the full model achieves an average ACC of 53.68% across seven datasets, outperforming the two variants by 5.63 and 5.07 percentage points, respectively. These comparisons provide direct evidence of the contributions of adaptive granular-ball representations and retrieved structural conditioning.

Variant/DatasetSynthetic3DEYaleB10COIL20MV100LeavesMfeatLandUse-21Scene-15
w/o C0.82660.21090.69440.33060.72850.22950.3821
Kmeans C0.79000.21560.61530.46120.67400.22520.3822
Ours0.82170.30000.77150.40560.80100.24430.4136

Furthermore, we extended the ablation study of training objectives in the main manuscript to missing rates of 0.1, 0.3, 0.5, and 0.7 across eight datasets. The results are presented in the following table. Among the 32 additional dataset–missing-rate settings, the full model achieves 19 individual best results and one tie. For example, at a missing rate of 0.7, the full model achieves an average ACC of 54.76%, outperforming the variants with only LGBL_{\mathrm{GB}} (39.09%) and only LdiffL_{\mathrm{diff}} (35.19%) by 15.67 and 19.57 percentage points, respectively. These results further support the complementary roles of region alignment and diffusion-based completion across different missing rates.

The benefits of combining the two components become more pronounced as the missing rate increases. In particular, removing LGBL_{\mathrm{GB}} leads to an average ACC reduction that increases from 4.72 percentage points at a missing rate of 0.1 to 19.57 percentage points at 0.7. This indicates that region alignment becomes increasingly important for diffusion-based completion as observational evidence becomes sparser. This finding is consistent with our mechanistic analysis: Evidence-Alignment promotes cross-view semantic alignment in the shared space through region correspondence, providing a basis for the query to retrieve semantically matched local prototypes from the target view. These prototypes, in turn, provide structural guidance for Structure-Aware Residual Diffusion. Overall, the results further demonstrate the complementary roles of the two components under highly incomplete settings.

Further component ablation studies and their corresponding results are provided in our response to the following comment. All additional ablation experiments reported in this response use seed 42. Due to the time and computational budget of this revision, the remaining ablations, diffusion formulation comparisons, and parameter analyses across additional missing rates and random seeds will be included in the appendix in a subsequent revision.

Missing Rate = 0.1

Variant/DatasetSynthetic3DEYaleB10COIL20MV100LeavesMfeatLandUse-21Scene-15Animal
w/o LGBL_{GB} & LdiffL_{diff}0.93660.24530.73400.74810.82800.26800.42630.2003
w/o LdiffL_{diff}0.97000.44060.75900.84250.77700.25330.46440.2285
w/o LGBL_{GB}0.96160.39060.75200.78560.80700.27950.44480.2200
ours0.97170.44370.83540.93060.85000.27480.48210.2300

Missing Rate = 0.3

Variant/DatasetSynthetic3DEYaleB10COIL20MV100LeavesMfeatLandUse-21Scene-15Animal
w/o LGBL_{GB} & LdiffL_{diff}0.83500.19840.54160.50120.54350.21090.32150.1721
w/o LdiffL_{diff}0.93160.37500.62290.63680.63300.22380.40640.1989
w/o LGBL_{GB}0.88830.33590.70970.60500.77550.24710.40570.2029
ours0.94170.42810.83060.88600.84050.28050.46110.2312

Missing Rate = 0.5

Variant/DatasetSynthetic3DEYaleB10COIL20MV100LeavesMfeatLandUse-21Scene-15Animal
w/o LGBL_{GB} & LdiffL_{diff}0.79500.19210.37500.34060.26050.15850.26370.1457
w/o LdiffL_{diff}0.92500.32030.53400.45560.48400.21380.34510.1626
w/o LGBL_{GB}0.72330.31710.63120.44750.70800.20520.37070.1898
ours0.92500.42500.83540.76690.83900.27330.45730.2089

Missing Rate = 0.7

Variant/DatasetSynthetic3DEYaleB10COIL20MV100LeavesMfeatLandUse-21Scene-15Animal
w/o LGBL_{GB} & LdiffL_{diff}0.72160.16250.29930.27180.20100.15000.21130.1201
w/o LdiffL_{diff}0.83500.31560.57700.34060.36000.22230.33080.1458
w/o LGBL_{GB}0.66000.22500.46250.30310.53650.16330.29540.1695
ours0.87830.37030.80830.61940.83300.27140.44190.1579
Reviewer_b1HtOpenReview ↗
Rating
4(+2)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

GRADE uses co-observed samples to align adaptive local regions across views, then retrieves target-view prototypes to guide residual diffusion for missing representation recovery. Experiments on eight datasets show overall improvements over the compared methods.

Critical strengths
  1. Using co-observed samples to establish soft correspondences between local regions is a reasonable design. It allows a region to match multiple regions in another view.

  2. I find the experimental evaluation fairly extensive.

Critical weaknesses
  1. Section 3 and Algorithm 1 use only the decrease in average distance to decide whether to split, with no conditions to prevent excessive splitting. This may produce singleton balls, reducing region-level matching to instance-level matching and limiting the aggregation of sparse cross-view evidence.

  2. The subtraction in Equation (13) does not ensure that the base contains only shared semantics or that the residual mainly captures target-view-specific local structure. The experiments do not analyze what each part learns.

  3. The ablations lack comparisons with fixed-granularity prototypes, no prototype conditioning, and observed-view conditioning. This makes it difficult to separate the contributions of adaptive partitioning, region alignment, and target-view structural guidance.

Questions for authors
  1. Apart from the decrease in DM, does Algorithm 1 use a minimum sample count, a splitting threshold, or any other stopping conditions?

  2. In Equation (6), two granular balls are treated as a valid negative correspondence if each has co-observation evidence but they share no members. When co-observed samples are scarce, the lack of overlap may simply reflect insufficient sampling. How do the authors avoid incorrectly treating semantically related regions as negative correspondences? Are the coverage weights in Equation (9) sufficient in this case?

  3. Section 5.1 constructs samples with either all views observed or only one view retained. Under the training conditions in Section 4.3, only complete samples can train the completion module. Their queries are formed by averaging the remaining (V-1) views, whereas queries during actual completion come from just one observed view. Could the authors explain this difference in more detail?

  4. Could the authors include more recent methods from 2026 in the comparisons to further validate the effectiveness of the proposed method?

6 comments

W1 & Q1:Apart from the decrease in DM, does Algorithm 1 use a minimum sample count, a splitting threshold, or any other stopping conditions?

We thank the reviewer for the careful inspection. We apologize for omitting two granularity-control parameters and their corresponding stopping criteria in Section 3 and Algorithm 1, which may have led the current description to be interpreted as continuously splitting granular balls solely based on decreases in weighted average distance. We will clarify these details in the revised manuscript.

Specifically, granular-ball construction consists of two stages. In the first stage, splitting is attempted only for parent balls containing at least nmin⁡=8n_{\min}=8 samples, and each of the two candidate child balls must contain at least two samples before the splitting-quality criterion is evaluated. In the second stage, a radius threshold is calculated based on the radii of the non-singleton balls obtained in the first stage:

Rdetect=min⁡{median(R),mean(R)}R_{\mathrm{detect}}=\min\lbrace\mathrm{median}(R),\mathrm{mean}(R)\rbrace,

Only balls with radii exceeding γRdetect\gamma R_{\mathrm{detect}} are further split, where γ=1.2\gamma=1.2. This detection threshold remains fixed throughout the second stage. Therefore, the actual construction process is jointly constrained by both the number of samples and the spatial scale, rather than continuing to split whenever the average distance decreases.

These conditions do not prohibit all singleton balls, but the algorithm does not require recursive splitting until every granular ball becomes a singleton. For granular balls containing multiple samples, cross-view correspondences are statistically established through the co-observation relationships among their members, allowing co-observation evidence to be aggregated at the region level rather than acting only on individual instances.

Furthermore, at a missing rate of 0.9, we conducted a sensitivity analysis of the two control parameters on seven datasets. The table reports ACC results, while the other evaluation metrics exhibit similar trends. Specifically, we investigated nmin⁡∈{4,8,16,32,64}n_{\min}\in\lbrace 4,8,16,32,64\rbrace and γ∈{1.0,1.2,1.4,1.6,1.8,2.0}\gamma\in\lbrace1.0,1.2,1.4,1.6,1.8,2.0\rbrace, varying one parameter at a time while keeping the other fixed at its default value. The results show that variations in the radius coefficient have a relatively moderate impact on performance across most datasets. In contrast, setting the minimum sample-count threshold too high leads to noticeable ACC degradation on some datasets, possibly because a higher splitting threshold limits the ability to capture local structures in greater detail. We uniformly adopt nmin⁡=8n_{\min}=8 and γ=1.2\gamma=1.2 throughout this study, without selecting dataset-specific optimal configurations.

Table 1. Sensitivity analysis of nmin⁡n_{\min}

Dataset48163264
Synthetic3D0.83670.82170.82500.82670.8100
EYaleB100.37190.30000.25940.21880.1938
COIL20MV0.72710.77150.63060.43540.2729
100Leaves0.43440.40560.35380.30000.2619
Mfeat0.78200.80100.77950.72850.5840
LandUse-210.25190.24430.25380.23430.2176
Scene-150.40780.41360.39150.38930.3773

Table 2. Sensitivity analysis of γ\gamma

Dataset1.01.21.41.61.82.0
Synthetic3D0.83170.82170.83330.83330.82670.8300
EYaleB100.32030.30000.29530.30000.29690.3078
COIL20MV0.74930.77150.74860.70140.72780.7035
100Leaves0.42130.40560.39690.39560.39440.3944
Mfeat0.79750.80100.79700.78850.78600.7800
LandUse-210.25380.24430.25620.25810.25710.2605
Scene-150.40220.41360.41630.40600.40690.4036

W2:The subtraction in Equation (13) does not ensure that the base contains only shared semantics or that the residual mainly captures target-view-specific local structure. The experiments do not analyze what each part learns.

We thank the reviewer for highlighting the roles of the two components. Equation (13) defines the division of responsibilities between base prediction and residual compensation. Their semantic roles are determined by the information sources used by the two branches, rather than guaranteed by the subtraction operation itself. The base is generated from the representations of other observed views in the aligned shared space to provide an initial prediction of the target view. Residual diffusion, conditioned on retrieved local prototypes from the target view, then recovers the target representation not accounted for by this prediction. Therefore, we emphasize cross-view-information-supported prediction and target-view-structure-guided compensation, rather than strict disentanglement of the two types of information. We will clarify the corresponding descriptions in the revised manuscript.

W3:The ablations lack comparisons with fixed-granularity prototypes, no prototype conditioning, and observed-view conditioning. This makes it difficult to separate the contributions of adaptive partitioning, region alignment, and target-view structural guidance.

We thank the reviewer for the suggestion. At a missing rate of 0.9, we conducted three additional ablation experiments on seven datasets and report the ACC results:

1.Without prototype conditioning (w/o C): The retrieved structural condition C is removed from the diffusion model.

2.Fixed-granularity prototypes (K-means C): Target-view granular-ball prototypes are replaced with K-means centroids to compare adaptive granular balls with a fixed number of prototypes.

3.Observed-view conditioning (Observed-view C): The conditioning source is changed from target-view granular balls to granular balls from currently observed views.

As shown in the table, the full model achieves an average ACC of 53.68% across seven datasets, outperforming K-means C, Observed-view C, and w/o C by 5.63, 1.96, and 5.07 percentage points, respectively. Specifically, on COIL20MV, adaptive granular balls improve ACC from 61.53% to 77.15% over K-means prototypes, demonstrating the effectiveness of constructing region representations based on data distributions. On 100Leaves, switching the conditioning source from observed-view to target-view granular balls increases ACC from 32.75% to 40.56%, highlighting the value of target-view structural information for recovery. On EYaleB10, introducing retrieved structural conditioning improves ACC from 21.09% to 30.00%, further demonstrating its contribution to clustering performance. These comparisons provide experimental evidence for GRADE's design from three perspectives: prototype construction, conditioning source, and conditioning usage.

Variant/DatasetSynthetic3DEYaleB10COIL20MV100LeavesMfeatLandUse-21Scene-15
w/o C0.82660.21090.69440.33060.72850.22950.3821
Kmeans C0.79000.21560.61530.46120.67400.22520.3822
Observed-view C0.82660.30620.74790.32750.79250.23850.3814
Ours0.82170.30000.77150.40560.80100.24430.4136

Q2:In Equation (6), two granular balls are treated as a valid negative correspondence if each has co-observation evidence but they share no members. When co-observed samples are scarce, the lack of overlap may simply reflect insufficient sampling. How do the authors avoid incorrectly treating semantically related regions as negative correspondences? Are the coverage weights in Equation (9) sufficient in this case?

We thank the reviewer for raising this concern. Equation (6) distinguishes unknown relations from zero-intersection relations. If either granular ball lacks commonly observed members, the correspondence is treated as unknown and excluded from alignment. Zero intersection is treated as a region-level negative correspondence only when both sides have common observation evidence. Equation (9) further adjusts regional contributions according to evidence coverage.

These designs reduce false negatives caused by treating unknown relations as negative, but cannot eliminate them under finite sampling. Notably, coverage reflects the extent of observation, not correspondence confidence, and cannot rule out zero intersection between semantically related regions due to insufficient sampling. We will clarify this limitation to avoid interpreting zero intersection as definitive semantic inconsistency.

Q3:Section 5.1 constructs samples with either all views observed or only one view retained. Under the training conditions in Section 4.3, only complete samples can train the completion module. Their queries are formed by averaging the remaining (V-1) views, whereas queries during actual completion come from just one observed view. Could the authors explain this difference in more detail?

We thank the reviewer for the comment. Based on the visibility patterns of incomplete samples, common missing-view scenarios can be categorized into two types:

  1. Single-view retention: Some samples remain complete, while the remaining samples retain only one view.
  2. Random partial-view retention: Incomplete samples may retain between 1 and V−1V-1 views.

This paper adopts the first setting. At the same incomplete-sample ratio, compared with the second setting, which allows multiple views to be retained, it provides less observed information and evaluates a more challenging recovery task with only a single information source.

Our recovery training is designed for general missing-view patterns, requiring only that the target view and at least one other view are observed to provide supervision and conditioning, respectively. Only under the current protocol do samples satisfying these requirements happen to be complete. The query aggregates representations in the shared space, while region-level soft contrastive alignment promotes cross-view semantic consistency. Ideally, representations of the same sample from different views should be close in the shared space. Accordingly, the single-view representation should also be close to the mean of its multi-view representations, providing a basis for transfer across different numbers of observed views.

To further evaluate applicability to more general missing-view scenarios, we conducted additional experiments by randomly removing 1 to V−1V-1 views at an incomplete-sample ratio of 0.9. In this setting, partially observed samples with at least two visible views can also satisfy the recovery training requirements. As shown in the table, GRADE achieves the best ACC on five of the seven datasets and the second-best results on the remaining two. On EYaleB10 and 100Leaves, it outperforms the respective best competing methods by 9.06 and 12.50 percentage points, respectively, supporting the applicability of our training design to different missing-view scenarios.

Method/DatasetSynthetic3DEYaleB10COIL20MV100LeavesMfeatLandUse-21Scene-15
DCP0.40830.20470.09930.11690.1910.09670.3313
PMIMC0.66500.16880.32010.27120.47400.14240.2118
FreeCSL0.64670.25620.67710.40940.81150.22810.3369
GUITARNaN0.32030.73680.51810.77250.25710.3072
HSACC0.35700.16880.52850.28060.13500.11570.2807
IMDC-DIA0.87670.18440.54100.28250.61350.16290.2725
IMVC-NCD0.87830.32660.76740.53190.9220.22950.3893
OPTION0.88660.24530.71660.39620.73950.22000.3678
GLGC0.83160.20150.42560.18870.57500.14660.3063
MAGIC0.53000.25780.45970.18810.69850.22950.3801
OURS0.89500.41720.74720.65690.85050.27000.4406
Reviewer_szq1OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GRADE for IMVC. The method has two parts. The first part builds granular balls in each view and uses jointly observed samples to link balls across views, so that local priors can be retrieved from the target view. The second part predicts a base representation from the observed views and uses diffusion to generate only the residual, conditioned on the retrieved local priors. Experiments on eight datasets with five missing ratios show that GRADE performs better than eight recent IMVC methods.

Critical strengths
  1. The motivation is clear. Existing imputation methods mainly rely on cross-view shared semantics and ignore the local structure of the target view.
  2. Using jointly observed samples as evidence to connect local regions across views is a reasonable idea, and the residual formulation of recovering only what the base representation misses is simple and intuitive.
  3. The experiments cover many datasets and missing ratios, and code is provided.
Critical weaknesses
  1. The granular-ball construction lacks any mechanism that prevents over-splitting (page 4, lines 189 to 192; page 15, Algorithm 1, steps 18 to 23). The only acceptance condition is DM_w(GB_g) < DM(GB_g), and splitting is applied recursively until no further split is possible, with no minimum ball size, complexity penalty, or other granularity control. A simple counterexample is the four one-dimensional points {0, 1, 10, 11}: the first split into {0, 1} and {10, 11} is accepted, and each child is then split again into singletons since the weighted measure drops to zero. For generic continuous latent features, the procedure as written therefore degenerates toward singleton balls, in which case region-level evidence reduces to instance-level matching and top-K prototype retrieval reduces to nearest-neighbor conditioning.
  2. Comparison with the closest diffusion-based IMVC methods is insufficient. DCG (AAAI‘25) and SACDG (TKDE’26) are discussed in Section 2 but are not included as baselines, leaving IMVC-NCD as the only diffusion-based competitor. Several relevant works are also not cited or positioned, including DMVG (Wen et al., ICML’24) for diffusion-based missing-view generation in IMVC, IDCD (Chen et al., ICML’26), whose conditional diffusion with disentangled shared and view-specific factors has a motivation very close to the shared-semantics versus target-view-specific information argument of this paper, and IASR (IJCAI‘26). The residual diffusion idea itself has prior art such as RDDM (Liu et al., CVPR’24), and the paper should clarify what is new relative to it.
Questions for authors
  1. The visibility pattern seen in training differs from the one seen at inference (page 7, lines 357 to 360; page 6, lines 276 to 282). In the experimental protocol a sample is either fully observed or retains a single view, whereas diffusion training uses only samples whose target view is observed and at least one other view is observed. Under this protocol those are exactly the fully observed samples, so the query during training aggregates V minus 1 views, while at inference the query for an incomplete sample comes from a single view. Why is the training set constructed this way, and how does this train-test mismatch affect recovery, especially at high missing ratios?
  2. What happens if the granular-ball prototypes are replaced by K-means prototypes or by nearest-neighbor instances in the style of IMVC-NCD as the conditioning signal? This comparison seems essential to support the adaptive-granularity contribution, and the released code already appears to contain these alternative options.
6 comments

W1:The granular-ball construction lacks any mechanism that prevents over-splitting (page 4, lines 189 to 192; page 15, Algorithm 1, steps 18 to 23). The only acceptance condition is DM_w(GB_g) < DM(GB_g), and splitting is applied recursively until no further split is possible, with no minimum ball size, complexity penalty, or other granularity control. A simple counterexample is the four one-dimensional points {0, 1, 10, 11}: the first split into {0, 1} and {10, 11} is accepted, and each child is then split again into singletons since the weighted measure drops to zero. For generic continuous latent features, the procedure as written therefore degenerates toward singleton balls, in which case region-level evidence reduces to instance-level matching and top-K prototype retrieval reduces to nearest-neighbor conditioning.

We thank the reviewer for the careful inspection. We apologize for omitting two granularity-control parameters and their corresponding stopping criteria in the main text and Algorithm 1, which may have led the current description to be interpreted as continuously splitting balls solely based on the decrease in weighted average distance. We will supplement these details in the revised manuscript. Specifically, granular-ball construction consists of two stages. In the first stage, splitting is attempted only for parent balls containing at least nmin⁡=8n_{\min}=8 samples, and each of the two candidate child balls is required to contain at least two samples before the splitting-quality criterion is evaluated. In the second stage, a radius threshold is calculated based on the radii of the balls obtained in the first stage:

Rdetect=min⁡{median⁡(R),mean⁡(R)}R_{\mathrm{detect}}=\min\left\lbrace \operatorname{median}(R),\operatorname{mean}(R)\right\rbrace

Only balls with radii exceeding γRdetect\gamma R_{\mathrm{detect}} are further split, where γ=1.2\gamma=1.2. This detection threshold remains fixed throughout the second stage. Therefore, the actual construction process is jointly constrained by both the number of samples and the spatial scale.

For the counterexample {0,1,10,11}\lbrace0,1,10,11\rbrace, the parent ball contains only four samples and thus does not satisfy the splitting threshold in the first stage. Its center and radius are both 5.5. At this point, Rdetect=5.5R_{\mathrm{detect}}=5.5, and its radius does not exceed 1.2×5.5=6.61.2\times5.5=6.6. Consequently, no further splitting occurs in the second stage either, and the final result retains a single ball containing all four points. These conditions do not prohibit all singleton balls, but they prevent the recursive degeneration in this counterexample that would otherwise be driven solely by decreasing distribution scales.

In addition, we conducted a sensitivity analysis of the two control parameters on seven datasets under a missing rate of 0.9. The table reports the ACC results, while the other evaluation metrics exhibit similar trends. Specifically, we investigated nmin⁡∈{4,8,16,32,64}n_{\min}\in\lbrace 4,8,16,32,64\rbrace and γ∈{1.0,1.2,1.4,1.6,1.8,2.0}\gamma\in\lbrace 1.0,1.2,1.4,1.6,1.8,2.0\rbrace, varying one parameter at a time while keeping the other fixed at its default value. The results show that variations in the radius coefficient have a relatively moderate impact on performance across most datasets. In contrast, setting the minimum sample-count threshold too high leads to noticeable performance degradation on some datasets, possibly because a higher splitting threshold limits the ability to capture local structures in greater detail. We uniformly adopt nmin⁡=8n_{\min}=8 and γ=1.2\gamma=1.2 throughout this study, without selecting dataset-specific optimal configurations.

Table 1. Sensitivity analysis of nmin⁡n_{\min}

Dataset48163264
Synthetic3D0.83670.82170.82500.82670.8100
EYaleB100.37190.30000.25940.21880.1938
COIL20MV0.72710.77150.63060.43540.2729
100Leaves0.43440.40560.35380.30000.2619
Mfeat0.78200.80100.77950.72850.5840
LandUse-210.25190.24430.25380.23430.2176
Scene-150.40780.41360.39150.38930.3773

Table 2. Sensitivity analysis of γ\gamma

Dataset1.01.21.41.61.82.0
Synthetic3D0.83170.82170.83330.83330.82670.8300
EYaleB100.32030.30000.29530.30000.29690.3078
COIL20MV0.74930.77150.74860.70140.72780.7035
100Leaves0.42130.40560.39690.39560.39440.3944
Mfeat0.79750.80100.79700.78850.78600.7800
LandUse-210.25380.24430.25620.25810.25710.2605
Scene-150.40220.41360.41630.40600.40690.4036

W2:Comparison with the closest diffusion-based IMVC methods is insufficient. DCG (AAAI‘25) and SACDG (TKDE’26) are discussed in Section 2 but are not included as baselines, leaving IMVC-NCD as the only diffusion-based competitor. Several relevant works are also not cited or positioned, including DMVG (Wen et al., ICML’24) for diffusion-based missing-view generation in IMVC, IDCD (Chen et al., ICML’26), whose conditional diffusion with disentangled shared and view-specific factors has a motivation very close to the shared-semantics versus target-view-specific information argument of this paper, and IASR (IJCAI‘26). The residual diffusion idea itself has prior art such as RDDM (Liu et al., CVPR’24), and the paper should clarify what is new relative to it.

We thank the reviewer for the suggestion. DCG's public implementation supports only two views, unlike our setting with three or more views. Runnable code for IDCD, IASR, and SACDG is unavailable, while DMVG's code lacks a complete clustering evaluation pipeline. We therefore added MAGIC and GLGC under the same experimental protocol.

We do not claim residual diffusion itself as novel. RDDM models image degradation via dual residual/noise diffusion, whereas GRADE predicts a target-view base from observed views and uses retrieved local prototypes to guide residual recovery. Our contributions are co-observation-driven adaptive region alignment and target-view structure-guided residual compensation, distinguishing GRADE from IDCD's information-theoretic disentanglement of shared and view-specific factors.

As shown in the table, GRADE achieves the best results in 33 of 35 dataset–missing-rate settings and ties in one. Its average ACC is 62.37%, outperforming GLGC (40.37%) and MAGIC (43.33%) by 22.00 and 19.04 percentage points, respectively. At a missing rate of 0.9, GRADE outperforms both methods on all seven datasets. On COIL20MV and Mfeat, it achieves 76.25% and 78.83% ACC, respectively, improving over the stronger competitor by 43.77 and 26.78 percentage points.

Missing Rate = 0.1

Method/DatasetSynthetic3DEYaleB10COIL20MV100LeavesMfeatLandUse-21Scene-15
GLGC (26'CVPR)0.91260.19960.52470.36410.69320.27900.3450
MAGIC (26'ICML)0.79900.21840.61540.43160.84000.24400.4042
Ours0.96860.42810.81260.92950.84690.27470.4698

Missing Rate = 0.3

Method/DatasetSynthetic3DEYaleB10COIL20MV100LeavesMfeatLandUse-21Scene-15
GLGC (26'CVPR)0.89390.21750.48380.28500.65670.27210.3439
MAGIC (26'ICML)0.88830.21620.58430.31360.76810.25100.3828
Ours0.94430.42960.80940.88310.83770.27590.4605

Missing Rate = 0.5

Method/DatasetSynthetic3DEYaleB10COIL20MV100LeavesMfeatLandUse-21Scene-15
GLGC (26'CVPR)0.84600.18680.45290.25400.61840.27030.3358
MAGIC (26'ICML)0.79730.23870.54080.22030.72500.25260.3882
Ours0.91460.41750.81830.76580.83630.27030.4515

Missing Rate = 0.7

Method/DatasetSynthetic3DEYaleB10COIL20MV100LeavesMfeatLandUse-21Scene-15
GLGC (26'CVPR)0.78690.18000.31950.20700.52380.23040.3146
MAGIC (26'ICML)0.63230.21530.43480.18660.71350.22630.3809
Ours0.87950.36870.80800.63430.82910.27200.4396

Missing Rate = 0.9

Method/DatasetSynthetic3DEYaleB10COIL20MV100LeavesMfeatLandUse-21Scene-15
GLGC (26'CVPR)0.73700.19250.21800.16410.37320.15350.2926
MAGIC (26'ICML)0.59630.20430.32480.15880.52050.16790.2827
Ours0.83030.30370.76250.40950.78830.25260.4054

Q1:The visibility pattern seen in training differs from the one seen at inference (page 7, lines 357 to 360; page 6, lines 276 to 282). In the experimental protocol a sample is either fully observed or retains a single view, whereas diffusion training uses only samples whose target view is observed and at least one other view is observed. Under this protocol those are exactly the fully observed samples, so the query during training aggregates V minus 1 views, while at inference the query for an incomplete sample comes from a single view. Why is the training set constructed this way, and how does this train-test mismatch affect recovery, especially at high missing ratios?

We thank the reviewer for the comment. Existing missing-view protocols can be broadly categorized into two types based on the visibility patterns of incomplete samples:

  1. Single-view retention protocol: Some samples remain complete, while each remaining incomplete sample retains only one view.
  2. Random partial-view retention protocol: Each incomplete sample randomly misses 1 to V−1V-1 views.

Accordingly, our recovery training is designed for general missing-view patterns: the ground-truth target view is available to provide supervision, and at least one other observed view is required to construct the conditioning input, without requiring the sample to be complete. Under the current protocol, samples satisfying these conditions happen to be complete. The query is obtained by averaging the representations of all views after mapping them into a shared semantic space. Region-level soft contrastive alignment promotes cross-view semantic consistency. Ideally, representations of the same sample from different views should be mapped to nearby positions in the shared space. Therefore, the single-view shared representation should also be close to the mean of the multi-view shared representations, providing a basis for knowledge transfer across different visible views. Under high missing rates, actual recovery performance still depends on alignment quality and the availability of sufficient supervision.

To further evaluate the applicability of our method to general missing-view patterns, we conducted additional comparative experiments at an incomplete-sample ratio of 0.9, where 1 to V−1V-1 views were randomly removed. As shown in the table, GRADE achieves the best ACC on five of the seven datasets and the second-best results on the remaining two. In particular, its ACC reaches 41.72% on EYaleB10 and 65.69% on 100Leaves, outperforming the best competing methods by 9.06 and 12.50 percentage points, respectively. These results support the applicability of our recovery training design to more general missing-view patterns. The best results are highlighted in bold, while the second-best results are underlined. Note that GUITAR reports NaN on Synthetic3D.

Method/DatasetSynthetic3DEYaleB10COIL20MV100LeavesMfeatLandUse-21Scene-15
DCP0.40830.20470.09930.11690.1910.09670.3313
PMIMC0.66500.16880.32010.27120.47400.14240.2118
FreeCSL0.64670.25620.67710.40940.81150.22810.3369
GUITARNaN0.32030.73680.51810.77250.25710.3072
HSACC0.35700.16880.52850.28060.13500.11570.2807
IMDC-DIA0.87670.18440.54100.28250.61350.16290.2725
IMVC-NCD0.87830.32660.76740.53190.9220.22950.3893
OPTION0.88660.24530.71660.39620.73950.22000.3678
GLGC0.83160.20150.42560.18870.57500.14660.3063
MAGIC0.53000.25780.45970.18810.69850.22950.3801
OURS0.89500.41720.74720.65690.85050.27000.4406

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.