acceptodds
Under review as a conference paper at ICLR 2027

Known-Novel Relation Learning for Multi-view Novel Class Discovery

Abstract

Novel Class Discovery (NCD) aims to cluster unlabeled samples from novel classes by transferring knowledge from labeled known classes, where the known and novel label spaces are disjoint. Existing NCD methods have achieved promising progress in single-view scenarios, but they are less effective for real-world multi-view data, in which different views provide both consistent and complementary information. Although recent multi-view NCD methods exploit distributional correlations between known and novel classes and perform adaptive view fusion, the similarity relations between known and novel categories remain insufficiently explored. In fact, novel classes are not isolated from known classes in feature or semantic spaces; their relation patterns with respect to known categories can provide valuable structural references for novel-class partitioning. To address this issue, this paper proposes a Known-Novel Relation Learning framework for Multi-view Novel Class Discovery. Specifically, we first employ multi-view matrix factorization to learn view-specific latent structures coupled by a shared partition representation. By constraining its known-class part with ground-truth labels, discriminative supervision is injected into multi-view representation learning and adaptive view-weight estimation, enabling distributional knowledge transfer to novel classes. Furthermore, known-class prototypes are regarded as relation anchors to construct a known-novel relation matrix, which represents novel samples in the reference space defined by known categories. By associating sample-level anchor relations with novel-class partitioning, samples assigned to the same novel class are encouraged to preserve consistent relation patterns, thereby explicitly transferring relational knowledge to novel-class clustering. These components are integrated into a unified optimization framework with an efficient iterative solver. Experiments on multiple multi-view datasets demonstrate the effectiveness of the proposed method.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
multi-view clustering, novel class discovery, relation learning

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_C5AvOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses the problem of discovering new visual categories and proposes a framework for learning the relationship between known classes and new categories. The method combines label-guided multi-view matrix decomposition with relation modeling based on known class prototypes, and jointly utilizes distribution information and category relationship information for new category clustering. It has achieved good experimental results on multiple benchmark datasets.

Critical strengths
  1. To address the issue that the existing multi-view new category discovery methods do not adequately utilize the relationship between known categories and new categories, a targeted solution has been proposed.
  2. By using the prototype of the known category as the relationship anchor point and guiding the division of new categories through the consistency constraint of the relationship, an effective combination of distributed information and relationship information has been achieved.
Critical weaknesses
  1. The method requires the pre-determination of the number of new categories, and the applicability in cases where the number of categories is unknown has not been discussed.
  2. The performance improvement on some datasets is relatively limited. It is recommended to further analyze the role of the relationship learning module under different data features.
  3. The experiments lack the standard deviation reports from multiple runs. Adding statistical stability analysis will make the results more convincing.
Questions for authors
  1. When the feature correlation between the known class and the new category is weak, will the effect of the relationship learning module be affected?
  2. Can the experimental results under different random seeds be supplemented to further verify the stability of the performance improvement?
1 comment

W1. What if the number of novel classes is unknown?

Our experiments assume that the number of novel classes kuk_u is given. Appendix F describes a possible extension. We would first estimate the count from labeled probe classes, following Han et al. (ICCV 2019), and then use the estimate to set the number of novel clusters and their relation prototypes.

W2. Why are the improvements smaller on some datasets?

On Reuters, our ACC is 0.24 percentage points above the strongest baseline in Table 2. Table 3 shows how much the relation branch contributes by keeping known-class supervision in both variants. On the original splits, adding the branch improves ACC by 2.50 points on Mfeat, 25.78 on Cora, and 29.10 on Citeseer. On Mfeat, the ablation already reaches 95.60% ACC, leaving less room for improvement.

We also compared the gain from the relation branch across known-class subsets with lower and higher similarity while keeping the novel samples fixed (see Q1 below).

W3. Report variability across runs.

We ran the full model and the relation-branch ablation on the original class splits using the same five seeds for both variants. For each dataset, we fixed the full model’s hyperparameters to the values used in Table 3. The ablation sets λ1=λ2=0\lambda_1=\lambda_2=0 and retains known-class supervision.

ACC over five seeds (mean ± sample SD, %)

DatasetFull modelWithout the relation branch
ProteinFold36.13 ± 2.2931.41 ± 2.47
Mfeat98.10 ± 0.0095.60 ± 0.07
Cora81.04 ± 1.9955.26 ± 4.54
Wiki73.96 ± 0.5862.80 ± 0.04
Citeseer64.96 ± 0.5535.86 ± 0.90

Q1. Does the relation branch still help when known and novel classes are weakly related?

For each dataset, we kept the novel samples fixed and matched the numbers of known classes and labeled samples across conditions. We used ground-truth labels to compute class centroids for this diagnostic. For each candidate subset of known classes, we averaged the cosine similarity of each known–novel centroid pair across views. We then took the highest similarity within that subset for each novel class and averaged these values across novel classes. We compared the lowest- and highest-scoring subsets using the same full-model hyperparameters.

Paired ACC gain from adding the relation branch (five seeds, mean ± SD, percentage points)

DatasetLower cosineHigher cosineGain at lower similarityGain at higher similarity
ProteinFold0.0700.185+1.73 ± 3.72+2.88 ± 2.56
Mfeat0.0740.267+19.58 ± 0.08+8.96 ± 13.95
Cora-0.114-0.070+7.02 ± 7.68+1.57 ± 2.33
Wiki-0.0500.262-4.85 ± 0.86-1.15 ± 4.00
Citeseer-0.171-0.137+22.88 ± 8.00+13.08 ± 6.14

Under lower similarity, the branch improved mean ACC on four of the five datasets. The gain was larger in the higher-similarity condition on two datasets and smaller on three. On Wiki, the branch reduced mean ACC in both conditions. The subsets also differ in which known classes they contain, so these results describe how the branch behaved under the two choices of anchors. One possible explanation for the gains at lower similarity is that novel classes can still differ in their relation patterns across the full set of anchors.

Q2. Does the improvement hold across random seeds?

On each of the five datasets, the full model achieved higher mean ACC than the ablation over the five runs (see W3). The difference between the two means ranged from 2.50 percentage points on Mfeat to 29.10 on Citeseer.

Reviewer_ba8GOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a matrix-factorization framework for multi-view novel class discovery, combining label-guided distribution transfer with known–novel relation learning. Known-class prototypes serve as anchors for representing novel samples, while relation consistency guides their clustering alongside feature reconstruction.

Critical strengths

1.The problem is practically relevant. With multi-view data becoming increasingly common, extending novel class discovery to exploit multiple feature sets or modalities is a worthwhile research direction.

2.The comparisons cover both multi-view clustering and novel class discovery, including several methods published in recent years. This provides a reasonably broad assessment of the proposed approach.

3.The experimental results support the method’s effectiveness across different types of data. The ablations also provide evidence for the contributions of its main components.

Critical weaknesses

1.Some optimization settings remain unspecified, particularly the stopping criterion and maximum iteration count. The convergence curves are useful, but the actual settings are still needed for reproducibility.

2.The implementation of some ablation variants needs clarification. For example, when known-class supervision is removed, it is unclear whether known-class samples still participate in training and how the corresponding prototypes are constructed.

3.The adaptation of single-view NCD baselines could be documented more precisely. Feature concatenation and MLP projection are described, but the projection architecture and training configuration are not specified.

4.The choice of one prototype per known class deserves further discussion. A single prototype may not capture multiple structures within a category. Explaining this choice and discussing the feasibility of multiple prototypes would help clarify the scope of the design.

Questions for authors

1.Could the framework be extended to cases where some samples have missing views? Which components would need to be modified?

2.This method assumes prior knowledge of kuk_u. How can it be applied when the number of novel classes is unknown?

1 comment

Thank you for identifying the implementation details that needed clarification.

W1. Specify the stopping criterion and maximum iteration count.

Section 4 and Algorithm 1 specify the update order and stopping rule. We use at most 100 outer iterations. Starting at the end of the third iteration, we stop when ∣Ft−Ft−1∣/max⁡(∣Ft∣,ϵmach)<10−5|F_t-F_{t-1}|/\max(|F_t|,\epsilon_{\mathrm{mach}})<10^{-5}. Each update of Qv\mathbf{Q}^{v} uses 80 projected-gradient steps. Figure 4 in Section 5.6 shows objective curves for all eight datasets.

W2. How are the known samples and prototypes handled without supervision?

In w/o Sup., we retain the known samples and apply kk-means to their standardized, concatenated features with klk_l groups. The group indicators replace the known-label block, and the factorization learns the view-specific anchors WvAvEl\mathbf{W}^{v}\mathbf{A}^{v}\mathbf{E}_{l} from those groups. Novel-only uses the novel samples alone. Both relation ablations retain the known samples and their ground-truth labels, as described in Section 5.3.

W3. Clarify the adaptation of single-view baselines.

For OpenNCD and SNCD, we concatenate the view features and feed them into a three-layer MLP encoder, as described in Section 5.1.2. Appendix C specifies the OpenNCD setup. We standardize each view before concatenation and use layer widths D→512→512→32D\to512\to512\to32, where DD is the concatenated feature dimension. We train OpenNCD for 40 epochs with Adam, an initial learning rate of 2×10−32\times10^{-3}, and a batch size of 256.

W4. Why use one prototype per known class?

One prototype per known class reduces the number of anchors from nln_l to klk_l and gives each relation coordinate a direct class correspondence. For classes with multiple modes, we could partition the labeled samples within each class into groups shared across views and learn a prototype for each group. Section 3.3 discusses this extension.

Q1. Can the method handle missing views?

The current formulation assumes complete views. Appendix F discusses feature recovery before factorization. A direct extension would restrict the reconstruction and relation losses to observed sample–view pairs and estimate each view weight from the corresponding observations.

Q2. What if the novel-class count is unknown?

The current method takes kuk_u as input. Appendix F discusses estimating it from labeled probe classes, following Han et al. (ICCV 2019), and using the estimate to set the number of novel clusters and relation prototypes.

Reviewer_VwiTOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents a known–novel relation learning framework for multi-view novel class discovery. Unlike existing methods that mainly focus on single-view data, the proposed approach considers multiple views and explores the relationships between known and novel classes. It uses known-class prototypes as anchors to learn relational features and guide the clustering of novel classes. These steps are integrated into a matrix-factorization framework. Experiments on eight multi-view benchmarks show promising results.

Critical strengths
  • The paper is well written and easy to understand.

  • Learning the relationships between known and novel classes is an interesting idea. Using known-class prototypes as anchors helps transfer useful information to novel-class clustering.

  • The experiments are extensive and provide good support for the proposed method.

Critical weaknesses
  1. IICMVNCD also studies multi-view NCD using matrix factorization and adaptive view fusion. A clearer comparison with this method is needed to show the main differences and highlight the new contributions.

  2. The related work section is too general. More discussion of closely related multi-view clustering methods based on matrix factorization is needed to clarify what is new in this work.

  3. The introduction lacks practical examples of multi-view NCD. Adding some real-world applications would help explain the motivation and show why this problem is important.

Questions for authors
  1. Why are existing single-view NCD methods less effective for multi-view data? What challenges arise when using multiple views, and how does the proposed method address these challenges?

  2. Since novel-class clustering is unsupervised, how are the hyperparameters selected without using the true labels of novel classes? Please clarify the selection process used in the experiments.

  3. What are the main limitations of the proposed method? What possible improvements could be explored in future work?

1 comment

Thank you for the suggestions on the motivation and related work.

W1. How does the method differ from IICMVNCD?

Both methods use matrix factorization and adaptive view fusion. IICMVNCD learns a shared basis for known and novel samples within each view and uses known-label prediction loss to guide view fusion. We fix the known-label block of the shared partition. In each view, we reconstruct each novel sample as a convex combination of known-class prototypes and use the resulting coefficients to encourage consistent relation patterns within each novel cluster. Sections 1 and 2 clarify this difference. Table 2 reports higher ACC for our method on all eight datasets under the reported settings, and Table 3 reports the relation ablations.

W2. Discuss related matrix factorization methods.

Section 2 now covers consensus factorization, one-step clustering, adaptive view weighting, and recent methods that use manifold learning, one-hot label constraints, or pairwise supervision. In our setting, labels from known classes guide clustering in a disjoint set of novel classes.

W3 and Q1. Motivate the multi-view setting and explain the challenge for single-view methods.

Section 1 uses video action discovery as an example. Appearance and motion offer complementary cues for discovering unseen actions from labeled examples of known actions. Concatenation presents the views as a single feature vector, so their individual contributions remain implicit. We learn view-specific factors and adaptive reconstruction weights, and couple the views through a shared novel partition.

Q2. How are hyperparameters selected when novel labels are unavailable?

For the reported benchmark results, we select λ1\lambda_1 and λ2\lambda_2 by grid search using ACC on the target novel classes, as described in Section 5.1.3. This selection uses ground-truth novel labels. In each model fit, known-class labels constrain the labeled partition. When target novel labels are unavailable and there are enough labeled classes, one option is to hold out whole known classes as surrogate novel classes, tune on that validation split, and freeze the selected parameters before clustering the target samples.

Q3. What are the limitations and possible extensions?

The method assumes a given novel-class count and complete, aligned views. Appendix F discusses estimating the count from labeled probe classes, following Han et al. (ICCV 2019), and handling missing views through feature recovery or updates based on observed sample–view pairs. Section 3.3 discusses using multiple prototypes to capture within-class variation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.