acceptodds
Under review as a conference paper at ICLR 2027

SPORT: Structure-Aware Prototype Disentanglement for Incomplete Multi-View Clustering

Abstract

Prototype-based incomplete multi-view clustering (IMVC) has recently gained attention for leveraging prototypes to recover missing views. However, existing methods often overlook view-specific semantics in prototypes, fail to preserve cluster-level relational structures, and rely on global prototypes without exploiting local geometric information, resulting in limited multi-view expressiveness and inaccurate imputation. To address these limitations, we propose a novel framework termed **S**tructure-aware **P**r**O**totype disentanglement fo**R** incomplete multi-view clus**T**ering (SPORT), which explicitly disentangles shared and view-specific components of prototypes while preserving cluster-level relational structures. Specifically, we decouple prototypes into orthogonal shared and view-specific components, aligning only shared components to capture consensus semantics while decorrelating view-specific components to preserve complementary information. Meanwhile, a structure-aware contrastive learning mechanism is incorporated to explicitly model cluster-level relationships during cross-view representation learning. Furthermore, a hybrid imputation strategy integrates global prototype matching with local neighborhood matching, enabling joint exploitation of semantic prototypes and manifold structures for missing-view recovery. Extensive experiments on six benchmark datasets show that SPORT achieves superior performance over state-of-the-art methods under various missing rates.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-view clustering

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_jd1iOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses incomplete multi-view clustering by proposing SPORT, which aims to preserve cross-view shared information while retaining view-specific characteristics. Its main components include a structure-aware relational contrastive objective, a prototype disentanglement mechanism based on shared and view-specific subspaces, and a hybrid imputation strategy that combines prototype matching with local neighborhood information. Experiments on six datasets at different missing rates, together with baseline comparisons and ablation studies, are presented to demonstrate the effectiveness of the proposed framework.

Critical strengths

Rather than enforcing complete prototype alignment across views, SPORT explicitly separates shared and view-specific components. This is a meaningful design choice for heterogeneous multi-view data. Extensive experiments further demonstrate the effectiveness of the proposed method.

Critical weaknesses

1.Prototype-based estimates and neighborhood-based estimates may have different reliability depending on the sample and the observed views. Consequently, does the use of fixed mixing coefficients limit the adaptability of the imputation process?

2.The relational weights are derived from cross-view similarity information. When only a small number of views are jointly observed, these similarity estimates will become less reliable, potentially affecting the contrastive learning process.

3.How sensitive is SPORT to the mixing coefficient used to combine prototype and neighborhood information? Could the authors report results for a few representative coefficient values on one or two datasets?

4.When the number of jointly observed views is small, how does the method avoid unreliable similarity estimates in the relational contrastive objective?

5.In the convergence analysis, why is it not the case that the changes in the loss value are shown, but only the changes in performance are presented?

Questions for authors

See in Critical Weaknesses.

4 comments

We thank reviewer jd1i for the valuable comments. Please find our response below.

1. Sensitive analysis of the hybrid prototype imputation strategy.

In Appendix G Figure 8 we analyze the performance of SPORT with different coefficient γ on all six datasets with 0.5 missing rate. From Figure 8 we can see that the performance of SPORT is relatively stable when γ is set from 0 to 0.4. For example, as shown in Figure 8 (c) on the Digit4k dataset, when γ varies between 0 and 0.4, the largest fluctuation in ACC is only by 2.40%. Meanwhile as shown in Figure 8 (d), on the 100Leaves dataset, the largest fluctuation in ACC is 4.2% as γ is set between 0 and 0.4. These results validate the robustness of SPORT's imputation strategy.

2. Discussions of the fixed coefficient for prototype imputation.

We agree that the performance of SPORT may vary due to different dataset characteristics by fixing γ. However, as shown in Figure 8 and discussed above, SPORT demonstrate a competitive and stable performance when γ varies from 0 to 0.4. This is because of the hybrid imputation strategy effectively balances prototype-level global structure and neighborhood-level local consistency. By integrating these two complementary sources of information, our model achieves more reliable reconstruction of missing views, which in turn leads to improved feature quality and more robust clustering performance across different datasets.

3. Discussion of relational weights under limited jointly-observed views.

We agree that the reliability of relational weights may decrease when only a limited number of views are jointly observed. Our work promotes the reliability under limited jointly-observed views in two aspects. First, instead of using a threshold to determine positive pairs based on similarity weights, we use soft weights smaller than 1 to capture structural relationships, thereby causing a smaller influence when structural relationships are unreliable. Second, the cross-view features corresponding to the same sample always have a weight of one, ensuring that the contrastive learning objective retains its basic cross-view alignment function even when inter-sample relational weights are less reliable. We will preserve further improving the reliability of structural information in future works

4. Explanation of the convergence analysis.

We thank reviewer jd1i for suggesting that a reporting the loss curve would be more sufficient for our convergence analysis. Previously we only report the performance because SPORT adopts a two-staged optimization framework where we use different losses during the pretraining and fine-tuning phase. Therefore, plotting the loss curve would seem a little disorganized due to the discontinuity between epochs 200 and 201. We present the convergence behavior of SPORT on Digit4k and 100Leaves with 0.5 missing rate in the following tables.

Table 1: Training loss on Digit4k during pretraining with 0.5 missing rate.

EpochLossEpochLossEpochLossEpochLossEpochLoss
20.103420.033820.0121220.0061620.004
40.049440.032840.0121240.0061640.004
60.041460.031860.0111260.0061660.004
80.039480.029880.0111280.0061680.004
100.038500.027900.0111300.0061700.004
120.038520.026920.0101320.0061720.004
140.038540.024940.0101340.0051740.004
160.038560.022960.0101360.0051760.003
180.038580.021980.0091380.0051780.003
200.038600.0191000.0091400.0051800.003
220.038620.0181020.0091420.0051820.003
240.038640.0171040.0081440.0051840.003
260.038660.0161060.0081460.0051860.003
280.038680.0161080.0081480.0051880.003
300.038700.0151100.0081500.0051900.003
320.037720.0141120.0071520.0041920.003
340.036740.0141140.0071540.0041940.003
360.036760.0131160.0071560.0041960.003
380.035780.0131180.0071580.0041980.003
400.034800.0121200.0071600.0042000.003

Table 2: Training loss on Digit4k during fine-tuning with 0.5 missing rate.

EpochLossEpochLossEpochLossEpochLossEpochLoss
202591.593262283.146322204.786382203.208442202.876
204588.490264258.655324204.707384203.176444202.876
206585.559266243.328326204.632386203.143446202.881
208583.321268232.693328204.561388203.111448202.880
210581.521270225.369330204.492390203.079450202.880
212579.750272220.213332204.426392203.047452202.879
214578.309274216.583334204.362394203.013454202.880
216576.831276213.988336204.301396202.979456202.880
218575.392278212.109338204.241398202.950458202.879
220573.933280210.727340204.183400202.930460202.881
222572.445282209.694342204.126402202.910462202.881
224570.899284208.907344204.071404202.892464202.881
226569.276286208.296346204.016406202.880466202.882
228567.570288207.810348203.962408202.878468202.880
230565.709290207.415350203.909410202.882470202.880
232563.672292207.087352203.856412202.881472202.880
234561.445294206.808354203.803414202.880474202.880
236558.944296206.565356203.751416202.877476202.879
238556.069298206.350358203.698418202.878478202.880
240552.676300206.158360203.647420202.880480202.879
242548.562302205.982362203.594422202.877482202.879
244543.327304205.820364203.542424202.877484202.879
246536.343306205.670366203.490426202.875486202.879
248526.511308205.530368203.442428202.876488202.878
250511.906310205.399370203.407430202.876490202.879
252489.314312205.274372203.373432202.875492202.879
254455.026314205.157374203.338434202.876494202.878
256410.483316205.052376203.305436202.877496202.880
258366.153318204.958378203.272438202.876498202.879
260323.197320204.869380203.241440202.876500202.878

Table 3: Training loss on 100Leaves during pretraining with 0.5 missing rate.

EpochLossEpochLossEpochLossEpochLossEpochLoss
20.122420.014820.0071220.0051620.004
40.052440.013840.0071240.0051640.004
60.052460.013860.0071260.0051660.004
80.048480.012880.0071280.0051680.004
100.045500.012900.0071300.0051700.004
120.043520.012920.0071320.0051720.004
140.037540.011940.0061340.0051740.004
160.035560.011960.0061360.0051760.004
180.032580.011980.0061380.0051780.004
200.030600.0101000.0061400.0051800.004
220.029620.0101020.0061420.0051820.004
240.027640.0091040.0061440.0051840.004
260.025660.0091060.0061460.0051860.004
280.023680.0091080.0061480.0051880.004
300.021700.0091100.0061500.0051900.004
320.019720.0081120.0061520.0051920.004
340.018740.0081140.0061540.0051940.004
360.017760.0081160.0061560.0051960.004
380.016780.0071180.0051580.0041980.004
400.015800.0071200.0051600.0042000.004

Table 4: Training loss on 100Leaves during fine-tuning with 0.5 missing rate.

EpochLossEpochLossEpochLossEpochLossEpochLoss
202117.269262108.835322108.323382108.246442108.231
204116.416264108.792324108.314384108.244444108.232
206115.616266108.755326108.307386108.244446108.235
208114.884268108.729328108.300388108.244448108.234
210114.228270108.696330108.296390108.242450108.239
212113.651272108.666332108.290392108.240452108.234
214113.171274108.639334108.291394108.242454108.236
216112.740276108.610336108.279396108.240456108.234
218112.354278108.593338108.280398108.238458108.237
220112.010280108.565340108.281400108.237460108.236
222111.688282108.554342108.277402108.238462108.234
224111.403284108.532344108.276404108.244464108.236
226111.130286108.523346108.270406108.240466108.234
228110.880288108.496348108.266408108.233468108.234
230110.646290108.483350108.262410108.231470108.232
232110.426292108.464352108.260412108.232472108.240
234110.238294108.450354108.259414108.236474108.240
236110.072296108.433356108.258416108.233476108.239
238109.917298108.423358108.252418108.234478108.242
240109.771300108.404360108.252420108.234480108.235
242109.636302108.398362108.254422108.236482108.231
244109.512304108.381364108.254424108.233484108.234
246109.401306108.373366108.251426108.236486108.233
248109.301308108.362368108.250428108.233488108.234
250109.208310108.349370108.246430108.232490108.228
252109.125312108.342372108.253432108.235492108.234
254109.056314108.336374108.244434108.234494108.232
256108.990316108.334376108.245436108.231496108.232
258108.932318108.324378108.248438108.232498108.235
260108.880320108.323380108.245440108.230500108.232

From the above tables we can see that the proposed framework reaches stable convergence within a moderate number of epochs, and further training yields negligible improvements.

Thank you for the author's response; my questions have been resolved.

Reviewer_mY9bOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a structure-aware prototype disentanglement framework for incomplete multi-view clustering, named SPORT. Specifically, SPORT first decomposes view-specific prototypes into shared and complementary components. After that, SPORT learns consistency for the shared components, while decorrelating the complementary ones. Then, SPORT proposes a similarity-weighted contrastive learning strategy and recovers missing views by integrating prototype matching with neighbor matching. Experiments on multi-view benchmarks demonstrate competitive performance of SPORT compared with existing SOTA methods.

Critical strengths
  1. The paper addresses an important problem in incomplete multi-view clustering, namely preserving view-specific information during prototype alignment.

  2. Experiments on various datasets demonstrate the effectiveness of the proposed SPORT.

  3. The paper is well structured and supported by extensive ablation studies.

Critical weaknesses
  1. My main concern is the motivation of this work. Although the authors claim three limitations of existing prototype-based IMVC methods in Introduction, these issues have already been explored in previous studies. Specifically, preserving view-specific information is one of the keys in multi-view clustering, which has been explored in many IMVC methods. Besides, employing neighborhood information to improve performance in clustering and classification tasks is widely-used strategy, e.g., Learning to Classify Images without Labels. As for preserving local neighborhood relationships during the recovering imcomplete information, I believe that this has also been discussed in related work on graph-based missing-view recovery, e.g., Decoupled Contrastive Multi-View Clustering with High-Order Random Walks.

  2. As discussed in Weakness 1, these limitations might remain insufficiently addressed in existing prototype-based IMVC methods, but they have been investigated and addressed in other IMVC methods. As a result, I agree that the novelty of this paper is questionable. The authors should clarify the differences between SPORT and other IMVC methods, rather than limiting the discussion to prototype-based methods.

  3. Besides, I am also concerned about the adequacy of the experiments. The authors devote much effort to discussing the limitations of existing prototype-based IMVC methods, but the experiments do not validate the corresponding claims and thus make the claims less convincing. For example, the paper lacks direct evidence verifying how SPORT preserves view-specific information or local neighborhood relationships, particularly compared with other IMVC methods. In other words, the authors should provide more experiments to verify that the proposed three modules truly resolves the three limitations.

  4. The font in Fig. 1 is too small, and the figure does not clearly demonstrate the rationale behind the method. The textual explanations could be simplified and replaced with clearer graphical illustrations.

Questions for authors

See the weaknesses.

5 comments

We thank reviewer mY9b for the valuable comments. Please find our response below.

1. Comparison with existing IMVC methods.

We agree that ideas such that employing neighborhood information to improve performance in clustering is not new to IMVC literature. However, our work is not a combination of previous work but can be seen as a complementary unified framework that jointly addresses the mentioned problems using prototype-based learning paradigm. We compare our SPORT with representative IMVC methods in the following.

Preserving view-specific information.

We agree that preserving view-specific information is a key factor of multi-view clustering and has been investigated in previous literature [1-4]. However, our method differs from [1-4] for that SPORT focus on preserving the complementary roles of cross-view consistency and view specific information from a prototype-level perspective which [1-4] do not focus on. Specifically, SPORT uses prototypes as representative semantic anchors of each view and then decomposes each prototype into shared and view-specific components, aligning only the shared components to capture cross-view consistency while encouraging the view-specific components to preserve complementary information through decorrelation. In contrast, [1] only considered learning view-specific prototypes while neglecting cross-view consistency. [2] focused on view-specific information from the feature-level and used dual-path autoencoders to decompose features into view-common and view-specific embeddings. [3,4] also focused on view-specific information from the feature-level similar to [2]. Therefore, our focus on preserving view-specific information differs from previous works.

Employing neighborhood information to improve performance.

In our work, we adopt contrastive learning to refine the quality of the learned features while leveraging neighborhood information to improve performance. From using similarity weights, SPORT is able to softly encode the neighboring information of multi-view data. [5] mined the nearest neighbors of each image based on feature similarity. However, they only focus on single view while we investigated the alignment problem of multi-view data. [9] used a contrastive-free approach by using GCN message passing and graph propagation to refine representations. Their work and ours adopted two parallel popular representation learning paradigms: contrastive learning and graph-based learning. [6-8] are close related works where they leveraged neighboring information to construct positive pairs beyond features from the same sample across views. However, their construction resulted in hard positive pairs which can be viewed as setting our similarity weight to 1, this may lead to a performance degradation when there are limited jointly-observed views.

Leveraging local neighborhood relationships during recovery.

We agree that leveraging local neighborhood relationships during recovery is widely adopted in previous works [10-14]. However, the novelty of our work lies in that it uses a hybrid prototype imputation strategy which jointly captures cluster-level and structural-level information for missing view recovery. On the contrary, [10-14] recovered missing data based on mutua information or graphs, which only exploits the neighboring information of multi-view data.

2. Discussion of the Motivation of SPORT.

As discussed in response 1, although the three issues have been investigated in previous IMVC studies, SPORT addresses them from a unified prototype-based perspective. Specifically, prototypes serve as semantic anchors that capture both cross-view consistency and view-specific information, while structure-aware contrastive learning preserves inter-sample relationships. Furthermore, hybrid prototype imputation combines global cluster-level semantics with local neighborhood information for missing-view recovery. Rather than treating these objectives separately, SPORT integrates them into a unified framework to learn more discriminative representations and improve missing-view recovery.

3. Additional experiments and analysis for SPORT.

We provide additional experiments to validate the effectiveness of our proposed methods.

Cross-view consistency of shared prototypes. We justify that the shared components capture cross-view consistency of prototype representations in the following experiment. We use samples that are observed in all views and assign each feature of these samples to its nearest prototype based on cosine similarity. Following the decomposition in Eq. (6), we separately perform this assignment using (i) shared components of prototypes, (ii) view-specific components of prototypes, and (iii) complete prototypes. We measure cross-view agreement as the percentage of samples assigned to the same prototype index across views. This agreement is averaged over all views if there exists more than two views. All results are reported as mean ±\pm standard deviation over three independent runs.

Table 1: Evaluation of cross-view consistency of shared prototypes on Digit4k and 100Leaves with different missing rate.

DatasetMissing RateShared Components Agreement (%)View-specific Components Agreement (%)Complete Prototypes Agreement (%)
Digit4k0.178.77±15.50\mathbf{78.77 \pm 15.50}66.46±14.9666.46 \pm 14.9673.75±15.0773.75 \pm 15.07
Digit4k0.376.26±12.63\mathbf{76.26 \pm 12.63}62.31±13.9362.31 \pm 13.9371.21±14.8271.21 \pm 14.82
Digit4k0.574.64±6.45\mathbf{74.64 \pm 6.45}60.73±9.2060.73 \pm 9.2068.87±9.7168.87 \pm 9.71
Digit4k0.767.11±7.32\mathbf{67.11 \pm 7.32}59.61±6.5959.61 \pm 6.5962.39±7.3162.39 \pm 7.31
100Leaves0.128.88±0.91\mathbf{28.88 \pm 0.91}17.96±0.9317.96 \pm 0.9326.39±0.5326.39 \pm 0.53
100Leaves0.330.59±1.12\mathbf{30.59 \pm 1.12}19.26±0.8219.26 \pm 0.8223.64±0.9323.64 \pm 0.93
100Leaves0.527.74±1.10\mathbf{27.74 \pm 1.10}17.01±1.0217.01 \pm 1.0222.80±1.1322.80 \pm 1.13
100Leaves0.721.90±2.00\mathbf{21.90 \pm 2.00}10.97±1.5510.97 \pm 1.5515.36±1.4915.36 \pm 1.49

As shown in Table 1, the shared components consistently achieve higher cross-view assignment agreement than view-specific and complete prototypes across all missing rates on both datasets. Thereby justifying that the shared components capture cross-view consistency of prototype representations.

View-specific information in view-specific prototypes.

We further examine whether the view-specific components preserve information from their corresponding views. Following Eq. (6), each prototype is decomposed as pgv=sgv+cgv\mathrm{\mathbf{p}}_g^v = \mathrm{\mathbf{s}}_g^v + \mathrm{\mathbf{c}}_g^v. We compare the reconstruction quality of three inputs using the same decoder DvD_v: (i) the shared component sgv\mathrm{\mathbf{s}}_g^v, (ii) the complete prototype sgv+cgv\mathrm{\mathbf{s}}_g^v + \mathrm{\mathbf{c}}_g^v, and (iii) the shared component combined with a view-specific component from another view sgv+c~π(g)u\mathrm{\mathbf{s}}_g^v + \widetilde{\mathrm{\mathbf{c}}}_{\pi(g)}^u.

We assign evaluation samples to their nearest prototypes and compute the MSE between the reconstructed prototype and the mean input features of its assigned samples. Here, owv and otv denote own-view and other-view view-specific components, respectively.

Table 2: Prototype reconstruction MSE on Digit4k and 100Leaves at a missing rate of 0.5 (lower is better).

DatasetShared componentShared + owv componentShared + otv component
Digit4k1.47×10−1±5.15×10−31.47\times10^{-1} \pm 5.15\times10^{-3}1.20×10−3±3.41×10−4\mathbf{1.20\times10^{-3}} \pm 3.41\times10^{-4}1.07×10−1±9.56×10−31.07\times10^{-1} \pm 9.56\times10^{-3}
100Leaves9.38×10−3±7.37×10−49.38\times10^{-3} \pm 7.37\times10^{-4}1.04×10−3±1.46×10−4\mathbf{1.04\times10^{-3}} \pm 1.46\times10^{-4}8.99×10−3±7.86×10−48.99\times10^{-3} \pm 7.86\times10^{-4}

As shown in Table 2, adding the original view-specific component significantly reduces reconstruction error on both datasets. In contrast, replacing it with a view-specific component from another view leads to much higher reconstruction error. These results suggest that the view-specific components retain information that is useful for reconstructing their corresponding views.

Local neighborhood structure preservation.

To validate SPORT preserves local neighboring relationships, we measure neighborhood purity (Purity@5) on Digit4k and 100Leaves under different missing rates. Specifically, we identify each feature's five nearest neighbors using cosine similarity in the learned representation space and compute the proportion of neighbors sharing the same ground-truth label with the selected feature. We compare SPORT with six representative IMVC methods: COMPLETER (CVPR'21), ProImp (IJCAI'23), IMVC-IE (ICASSP'24), BURG (ICCV'25), BRIDGE (ICCV'25), and HSACC (NeurIPS'25). All results are reported as mean ±\pm standard deviation over three independent runs.

Table 3: Purity@5 on Digit4k under different missing rates.

Missing rateSPORTProImpCompleterIMVC-IEBURGHSACCBRIDGE
0.187.59±0.06\mathbf{87.59 \pm 0.06}86.72±0.2986.72 \pm 0.2985.06±0.4285.06 \pm 0.4282.84±0.8982.84 \pm 0.8987.32±0.0787.32 \pm 0.0785.70±0.2685.70 \pm 0.2645.10±0.0045.10 \pm 0.00
0.387.31±0.01\mathbf{87.31 \pm 0.01}86.74±0.2986.74 \pm 0.2986.05±0.2086.05 \pm 0.2077.08±2.7977.08 \pm 2.7986.30±0.1086.30 \pm 0.1084.31±0.5984.31 \pm 0.5964.46±0.0064.46 \pm 0.00
0.587.05±0.35\mathbf{87.05 \pm 0.35}86.45±0.3086.45 \pm 0.3084.78±0.5784.78 \pm 0.5766.26±3.4566.26 \pm 3.4585.92±0.1785.92 \pm 0.1784.56±1.1484.56 \pm 1.1461.99±0.0061.99 \pm 0.00
0.786.85±0.03\mathbf{86.85 \pm 0.03}86.24±0.4686.24 \pm 0.4685.46±0.3385.46 \pm 0.3366.28±7.7966.28 \pm 7.7985.95±0.2185.95 \pm 0.2184.67±0.6884.67 \pm 0.6862.11±0.0062.11 \pm 0.00

Table 4: Purity@5 on 100Leaves under different missing rates.

Missing rateSPORTProImpCompleterIMVC-IEBURGHSACCBRIDGE
0.189.88±0.39\mathbf{89.88 \pm 0.39}77.35±1.3677.35 \pm 1.3643.61±1.4643.61 \pm 1.4648.36±1.8848.36 \pm 1.8888.88±0.3688.88 \pm 0.3637.91±3.4037.91 \pm 3.4048.44±0.0048.44 \pm 0.00
0.380.24±0.73\mathbf{80.24 \pm 0.73}61.19±0.4561.19 \pm 0.4532.04±4.3632.04 \pm 4.3627.23±3.6627.23 \pm 3.6676.05±0.1976.05 \pm 0.1933.22±1.1633.22 \pm 1.1622.88±0.0022.88 \pm 0.00
0.570.47±1.20\mathbf{70.47 \pm 1.20}58.58±1.1358.58 \pm 1.1341.16±9.8441.16 \pm 9.8416.74±2.7116.74 \pm 2.7164.49±0.5364.49 \pm 0.5339.90±2.2639.90 \pm 2.2618.73±0.0018.73 \pm 0.00
0.755.73±0.75\mathbf{55.73 \pm 0.75}47.04±1.7447.04 \pm 1.7430.95±0.7330.95 \pm 0.738.30±2.748.30 \pm 2.7453.43±0.0853.43 \pm 0.0829.38±2.6829.38 \pm 2.6819.26±0.0019.26 \pm 0.00

From table 3 and 4 we can see that SPORT consistently achieves the highest Purity@5 across all eight settings. These results demonstrate that SPORT learns more semantically consistent local neighborhoods, supporting the effectiveness of its structure-aware relational contrastive learning.

In conclusion, the above demonstrates that SPORT resolves the three limitations mentioned in our work.

4. Adjustment for Figure 1.

We thank reviewer mY9b for suggesting a clearer figure would be more informative. We have uploaded the revised manuscript with a larger Figure 1.

Reviewer_RtUdOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a structure-aware prototype disentanglement framework for incomplete multi-view clustering (SPORT). The method combines relational contrastive learning with prototype partial alignment to separate shared and view-specific information. It further introduces a hybrid prototype imputation strategy that integrates prototype-level information and local neighborhood information to estimate missing views. The results suggest that the proposed achieves competitive clustering performance.

Critical strengths
  1. The integration of relational learning and prototype disentanglement is a strength of the proposed method for capturing shared and view-specific information.
  2. The hybrid imputation strategy effectively leverages both prototype-level and neighborhood-level information to estimate missing views. The results suggest that the proposed achieves competitive clustering performance.
Critical weaknesses
  1. The method uses an orthogonal projection to separate shared and view-specific components, but orthogonality alone does not necessarily guarantee that these components capture semantically shared and view-specific information, respectively. More evidence is needed to support the intended interpretation of this decomposition.
  2. It is not sufficiently clear whether the relational contrastive objective provides complementary information beyond prototype alignment or whether their effects are partially redundant. A clearer distinction from existing approaches should be made.
Questions for authors
  1. Could the authors provide an additional analysis to verify that the shared components preserve cross-view cluster consistency while the view-specific components retain view-dependent information
  2. How does the relational contrastive loss contribute beyond the prototype alignment loss?
  3. Prototype Over-alignment Problem (POP) issue requires a more clear explanation.
  4. Could you clarify the time cost incurred by the proposed approach.
3 comments

We thank reviewer RtUd for the valuable comments. Please find our response below.

1. Disentanglement of shared and view-specific prototype components.

Cross-view consistency of shared prototypes.

We justify that the shared components capture cross-view consistency of prototype representations through the following experiment. We use samples that are observed in all views and assign each feature of these samples to its nearest prototype based on cosine similarity. Following the decomposition in Eq. (6), we separately perform this assignment using (i) shared components of prototypes, (ii) view-specific components of prototypes, and (iii) complete prototypes. We measure cross-view agreement as the percentage of samples assigned to the same prototype index across views. This agreement is averaged over all view pairs when there are more than two views. All results are reported as mean ± standard deviation over three independent runs.

Table 1: Evaluation of cross-view consistency of shared prototypes on Digit4k and 100Leaves under different missing rates.

DatasetMissing RateShared Components Agreement (%)View-specific Components Agreement (%)Complete Prototypes Agreement (%)
Digit4k0.178.77 ± 15.5066.46 ± 14.9673.75 ± 15.07
Digit4k0.376.26 ± 12.6362.31 ± 13.9371.21 ± 14.82
Digit4k0.574.64 ± 6.4560.73 ± 9.2068.87 ± 9.71
Digit4k0.767.11 ± 7.3259.61 ± 6.5962.39 ± 7.31
100Leaves0.128.88 ± 0.9117.96 ± 0.9326.39 ± 0.53
100Leaves0.330.59 ± 1.1219.26 ± 0.8223.64 ± 0.93
100Leaves0.527.74 ± 1.1017.01 ± 1.0222.80 ± 1.13
100Leaves0.721.90 ± 2.0010.97 ± 1.5515.36 ± 1.49

As shown in Table 1, the shared components consistently achieve higher cross-view assignment agreement than view-specific components and complete prototypes across all missing rates on both datasets, thereby justifying that the shared components capture cross-view consistency of prototype representations.

View-specific information in view-specific prototypes.

We further examine whether the view-specific components preserve information from their corresponding views. Following Eq. (6), each prototype is decomposed as pgv=sgv+cgv\mathbf{p}_g^v = \mathbf{s}_g^v + \mathbf{c}_g^v. We compare the reconstruction quality of three inputs using the same decoder DvD_v: (i) the shared component sgv\mathbf{s}_g^v, (ii) the complete prototype sgv+cgv\mathbf{s}_g^v + \mathbf{c}_g^v, and (iii) the shared component combined with a view-specific component from another view sgv+c~π(g)u\mathbf{s}_g^v + \widetilde{\mathbf{c}}_{\pi(g)}^u.

We assign evaluation samples to their nearest prototypes and compute the MSE between the reconstructed prototype and the mean input features of its assigned samples. Here, owv and otv denote own-view and other-view view-specific components, respectively.

Table 2: Prototype reconstruction MSE on Digit4k and 100Leaves at a missing rate of 0.5 (lower is better).

DatasetShared ComponentShared + owv ComponentShared + otv Component
Digit4k1.47 × 10⁻¹ ± 5.15 × 10⁻³1.20 × 10⁻³ ± 3.41 × 10⁻⁴1.07 × 10⁻¹ ± 9.56 × 10⁻³
100Leaves9.38 × 10⁻³ ± 7.37 × 10⁻⁴1.04 × 10⁻³ ± 1.46 × 10⁻⁴8.99 × 10⁻³ ± 7.86 × 10⁻⁴

As shown in Table 2, adding the original view-specific component significantly reduces reconstruction error on both datasets. In contrast, replacing it with a view-specific component from another view leads to much higher reconstruction error. These results suggest that the view-specific components retain information that is useful for reconstructing their corresponding views.

We agree that our orthogonal projection approach may not be the perfect way to disentangle complementary information in IMVC. We leave this direction for future work.

2. Discussion of the effects of the relational contrastive loss and the prototype alignment loss.

The relational contrastive loss and the prototype alignment loss optimize different aspects of our framework. Specifically, the relational contrastive loss optimizes the complete features of samples across different views. By using pairwise similarity weights to characterize inter-instance affinities, the relational contrastive loss pulls the features of the same sample and features of neighbors closer in the latent space, thereby promoting cross-view consistency while preserving cluster-level relational structures.

In contrast, the prototype alignment loss optimizes the prototypes generated via k-means. By aligning the shared components while decorrelating view-specific components, the prototype alignment loss is able to capture consensus semantics while preserving complementary information.

As discussed in our ablation study in Table 2 of the paper, removing either of the two losses incurs a performance degradation. This design of using separate losses to optimize features and prototypes is also adopted in previous prototype-based works [1-5]. Therefore, the relational contrastive loss and the prototype alignment loss jointly contribute to our proposed framework.

3. Distinction from existing approaches.

We compare SPORT with representative prototype-based IMVC methods [1-5] in terms of three key components: (1) feature alignment, (2) prototype alignment, and (3) prototype imputation.

Feature alignment. Methods [1,3,5] adopt classic contrastive learning by only aligning the features of the same sample across views, which may push semantically similar samples apart in the latent space. Aware of such drawbacks, [4] designs a dynamic contrastive learning loss that dynamically selects the top δ\delta\\% most similar negative features as positives. In addition, [2] proposes a contrastive learning-free alignment loss using the ℓ2\ell_2 norm.

In our work, we move beyond classic contrastive learning by using similarity weights to softly encode structural relationships in the learning process via relational contrastive learning, thereby allowing semantically similar neighbors to also be pulled closer in the latent space. Our work can be seen as complementary to [4], as we develop a different strategy to mitigate the drawbacks of classic contrastive learning.

Prototype alignment. Methods [1-5] focus on building prototype correspondence across different views. Such overly strict alignment suppresses view-specific information, limiting the expressive capacity of multi-view representations, which we refer to as POP. Our work addresses this issue by decomposing prototypes into shared and view-specific components before alignment, thereby preserving both cross-view consistency and complementary view-specific information.

Prototype imputation. Methods [1-4] use either matched prototypes or neighbors to impute missing information based on cosine similarity. Method [5] imputes the missing features by aggregating features from the same sample and prototypes across views.

In contrast, SPORT adopts a hybrid prototype imputation strategy that jointly utilizes the structure-level and cluster-level information conveyed by neighbors and prototypes. Our hybrid prototype imputation strategy achieves better results than both neighbor-only and prototype-only imputation strategies, as discussed in Appendix G.

In conclusion, our approach is distinct from previous approaches in terms of feature alignment, prototype alignment, and prototype imputation.

4. Explanation of the Prototype Over-alignment Problem.

In prototype-based IMVC, prototypes serve as semantic anchors for each view, capturing representative cluster-level information and guiding missing-view imputation. Although corresponding prototypes across different views should share consistent cluster-level semantics, they also need to preserve view-specific characteristics arising from different observation processes. Enforcing overly strict alignment between entire prototypes may suppress these view-specific characteristics, resulting in the loss of complementary information across views and limiting the representational capacity of prototypes. We refer to this phenomenon as the Prototype Over-alignment Problem (POP). To address this issue, SPORT decomposes each prototype into shared and view-specific components, aligning only the shared components to capture cross-view consistency while encouraging the view-specific components to preserve complementary information through decorrelation.

5. Computation time of SPORT and comparison with representative works.

We report the computation time comparison between SPORT and representative baselines in the following tables. For a fair comparison, we report the mean and standard deviation of the total time required for model training and inference on Digit4k, 100Leaves, and ALoi100 under four missing rates over 10 independent runs.

Table 3: Comparison of computation time (seconds) on Digit4k.

MethodMR=0.1MR=0.3MR=0.5MR=0.7
ProImp52.99 ± 1.7942.88 ± 1.2333.47 ± 0.8624.43 ± 0.55
IMVC-IE189.70 ± 2.37213.15 ± 5.42209.43 ± 13.60161.28 ± 3.18
RPCIC513.81 ± 20.03471.35 ± 18.30428.83 ± 15.67387.81 ± 17.94
BURG132.51 ± 6.99132.50 ± 5.92135.76 ± 1.27133.13 ± 7.25
Completer89.44 ± 1.4881.86 ± 1.2153.54 ± 2.1236.69 ± 2.03
APADC348.42 ± 19.57371.37 ± 5.68364.19 ± 21.12363.54 ± 22.24
HSACC153.63 ± 4.43144.64 ± 1.80115.61 ± 1.3875.01 ± 1.81
IMC-MCL425.52 ± 6.85365.55 ± 4.51250.23 ± 1.21161.62 ± 1.63
Ours43.69 ± 0.2171.91 ± 0.8772.94 ± 0.6781.22 ± 0.84

Table 4: Comparison of computation time (seconds) on 100Leaves.

MethodMR=0.1MR=0.3MR=0.5MR=0.7
ProImp29.21 ± 0.3527.05 ± 0.1619.22 ± 0.3015.80 ± 0.27
IMVC-IE82.15 ± 1.7874.93 ± 0.6665.59 ± 1.7354.98 ± 0.57
RPCIC336.09 ± 16.41302.73 ± 14.88278.97 ± 3.12249.70 ± 0.56
BURG90.41 ± 4.1190.78 ± 4.8494.30 ± 1.1293.85 ± 1.16
Completer44.27 ± 2.1338.70 ± 2.0933.79 ± 0.5920.49 ± 0.33
APADC181.03 ± 12.05229.77 ± 14.03237.19 ± 1.24238.32 ± 1.62
HSACC66.58 ± 2.4857.34 ± 2.1548.02 ± 0.8638.39 ± 0.45
IMC-MCL221.12 ± 1.40185.14 ± 0.94148.19 ± 0.9086.11 ± 0.96
Ours27.47 ± 0.7338.13 ± 1.0139.42 ± 1.0343.26 ± 1.21

Table 5: Comparison of computation time (seconds) on ALoi100.

MethodMR=0.1MR=0.3MR=0.5MR=0.7
ProImp149.50 ± 3.68116.44 ± 0.3994.35 ± 2.8066.35 ± 1.69
IMVC-IE795.77 ± 16.711104.65 ± 35.621207.31 ± 45.47990.92 ± 28.49
RPCIC3291.49 ± 278.412962.71 ± 161.452634.59 ± 208.582337.03 ± 183.35
BURG799.05 ± 83.67867.67 ± 49.30869.59 ± 62.04857.50 ± 77.49
Completer234.32 ± 21.29205.03 ± 9.67147.82 ± 12.1491.76 ± 8.35
APADC386.86 ± 31.12415.57 ± 29.11416.27 ± 30.27414.79 ± 34.38
HSACC398.44 ± 17.58385.46 ± 18.45314.20 ± 13.89214.42 ± 10.70
IMC-MCL1492.40 ± 9.941186.23 ± 19.83904.73 ± 11.32594.40 ± 11.27
Ours565.36 ± 27.011250.15 ± 64.791489.86 ± 36.791519.39 ± 95.80

References

[1] Honglin Yuan, Yuan Sun, Fei Zhou, Jing Wen, Shihua Yuan, Xiaojian You, and Zhenwen Ren. Prototype matching learning for incomplete multi-view clustering. IEEE Transactions on Image Processing, 34:828–841, 2025.

[2] Jiaqi Jin, Siwei Wang, Zhibin Dong, Xinwang Liu, and En Zhu. Deep incomplete multi-view clustering with cross-view partial sample and prototype alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11600–11609, 2023.

[3] Honglin Yuan, Shiyun Lai, Xingfeng Li, Jian Dai, Yuan Sun, and Zhenwen Ren. Robust prototype completion for incomplete multi-view clustering. In Proceedings of the 32nd ACM International Conference on Multimedia, pp. 10402–10411, 2024.

[4] Zhengzhong Zhu, Pei Zhou, Lanxi Bai, Jia Nie, Li Cheng, Shiquan Min, and Jiangping Zhu. Prototype and sample level semantic alignment for incomplete multi-view clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5818–5827, 2026.

[5] Haobin Li, Yunfan Li, Mouxing Yang, Peng Hu, Dezhong Peng, and Xi Peng. Incomplete multiview clustering via prototype-based imputation. arXiv preprint arXiv:2301.11045, 2023.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.