Neighbor Contrast Learning Helps Clustering Cope With Missing Data Views

Real-world data rarely arrives complete. A single object—a tumor, a handwritten digit, a plant leaf—can be described by several kinds of measurements at once, and in practice some of those measurements are simply absent. A patient may never receive the CT scan that would have completed an imaging record; a sensor may fail before logging one channel of information. Datasets of this kind are called incomplete multi-view data, and clustering them—grouping unlabeled samples into meaningful categories when some views are missing for some samples—has become a central problem in data mining. A research team now reports a new method, called NCMS, that tackles the problem by leaning on a deceptively simple idea: even when a sample is missing from one view, its neighbors in the views that remain can tell you a great deal about it.

The method, described as a neighbor contrast-based multiple self-supervised approach to incomplete multi-view clustering, was developed by Yiwei Yu, Lihua Zhou, Guowang Du, Jialong Wang, Chao Liu, Ziyi Yin and colleagues, and published in the journal Data Science and Engineering. Its three modules work in sequence. First, missing views are filled in by transferring neighbor relationships across views: the model computes which samples are most similar to each other in the views that are observed, then uses the attribute information of those neighbors to estimate what the missing view would have contained. Second, a two-level representation learning strategy—contrastive learning within each view and across the fused global representation—extracts compact features from the completed data. Third, a multiple self-supervised learning module, built on KL divergence, uses soft cluster assignments generated during training as supervisory signals to refine the representation that ultimately drives the clustering.

The neighbor-transfer step rests on an assumption the authors trace to prior work: the same sample should share similar neighbors in different views, and samples should be feature-similar to their neighbors. If a sample is absent from a view, its features there can be inferred from the neighbors it has in other views, weighted by similarity. Samples that are themselves incomplete are removed from the neighbor set before filling, to keep the transferred information clean. The authors argue this avoids the pitfalls of cruder strategies such as zero padding or filling with the mean of all observed samples, which cannot reflect the true distribution of samples in each view, and of generative adversarial networks, which demand consistent data distributions and substantial computational resources and can suffer from unstable training.

The contrastive learning component departs from the standard recipe. Conventional contrastive methods create positive pairs by augmenting a sample—cropping or rotating an image, for instance—and draw negatives at random from the rest of the dataset. Randomly chosen negatives, however, may share features with the anchor sample and mislead the model. NCMS instead treats each sample’s identified neighbors as its positives and all non-neighbors as negatives, both within each view and in the fused global representation produced by a multi-channel attention mechanism. The authors argue that this shift from instance-level to neighborhood-level contrast provides a more robust supervision signal for incomplete data, because the collective representation of a sample’s semantically similar neighbors is more reliable than the potentially corrupted representation of the incomplete sample itself. They also formulate a contrastive loss that naturally accommodates multiple positive pairs per anchor, an extension they describe as non-trivial.

The self-supervised module addresses a weakness the authors identify in earlier pseudo-label approaches, which were largely designed for single-view data. In multi-view settings, the same sample may be assigned to different clusters in different views, producing conflicting guidance signals. NCMS generates soft cluster assignments separately for each view and for the global representation, using a Student-t distribution kernel over cluster centers initialized by k-means, then minimizes the KL divergence between the soft assignments and sharpened target distributions. The view-specific and global-specific losses are unified in a single optimization objective, so the model retains view-specific information while converging toward a shared clustering structure. Training proceeds in two phases: pre-training combines a scaled cosine similarity reconstruction loss with the two contrastive losses, and fine-tuning optimizes the multiple self-supervised loss before final cluster labels are read out from the global assignment matrix.

The experimental evidence comes from four widely used multi-view datasets: BDGP, a fruit fly embryo dataset with 2,500 samples in five categories and two views of visual and textual features; Caltech101-20, 2,386 images across 20 categories with HOG and GIST features; 100Leaves, 1,600 samples covering 100 plant leaf types with three views of shape, edge-histogram and texture descriptors; and Handwritten, 2,000 images of ten digits with five feature views. Missingness was simulated by randomly sampling a binary availability mask at target missing rates, with each sample guaranteed to remain observable in at least one view, and the same masks were used for all compared methods under each of ten random seeds. Nine baselines were evaluated, including simple strategies such as best single-view clustering and feature concatenation, traditional matrix factorization, subspace and graph-based methods, and deep approaches including COMPLETER, AGCL and PMIMC.

The headline result concerns 100Leaves. When half the view entries were missing, NCMS improved clustering accuracy by 6.16 percent and normalized mutual information by 4.44 percent over the next-best method. More broadly, NCMS achieved the best performance under all missing rates on the two datasets with more than two views, 100Leaves and Handwritten, which the authors attribute to richer neighbor information supporting higher-quality completion. Performance of all methods generally deteriorated as missing rates rose, but NCMS remained competitive in the hardest settings: on BDGP at a 90 percent missing rate it recorded an accuracy of 84.03 plus or minus 1.14, against 81.28 plus or minus 2.36 for AGCL, while PMIMC took the top accuracy and NMI on that dataset at that extreme rate. AGCL also edged out NCMS on some Caltech101-20 NMI values and on BDGP accuracy at a 70 percent missing rate.

The authors are notably candid about where their method does not win. Wilcoxon signed-rank tests against PMIMC, run over ten matched seed pairs per dataset and missing rate, showed statistically significant improvements for NCMS in most configurations—consistently so on Handwritten across all missing rates—but also identified settings where NCMS was not significantly better and could be slightly worse, notably on BDGP and 100Leaves at high missing rates. Confidence intervals over the ten runs told a similar story: clear separation from PMIMC in most settings, but gaps that narrowed or reversed on BDGP at 90 percent missing and on 100Leaves at 50 percent, where PMIMC’s intervals lay entirely above NCMS’s. The overall pattern, the authors conclude, is that NCMS performs better at low missing rates and in multi-view settings, while PMIMC gradually gains the advantage as missingness increases—an outcome they flag as an area needing improvement.

Ablation experiments at a 50 percent missing rate showed that every module contributes: removing the reconstruction loss hurt most, followed by removing view-specific or global-specific contrastive learning, with the multiple self-supervision module contributing least. Sensitivity analyses found the optimal neighbor number, loss weights and batch size vary by dataset and missing rate, and the method’s time complexity, quadratic in the number of samples per view pair, limits its use on large-scale data. The authors also acknowledge that pseudo-label quality can suffer under class imbalance, that the fixed number of neighbor positives cannot adapt to model confidence, and that the method assumes views contain complementary yet consistent information, making it unsuitable for inherently inconsistent views. Beyond clustering, they suggest the framework could inspire solutions for other incomplete multi-view tasks such as classification, and the code is publicly available on Gitee.

Subject of Research: A self-supervised machine learning method for clustering data with missing views using neighbor contrastive learning

Article Title: Neighbor Contrast-based Multiple Self-supervised Incomplete Multi-view Clustering

Article References: Yu, Y., Zhou, L., Du, G., Wang, J., Liu, C., & Yin, Z. (2026). Neighbor Contrast-based Multiple Self-supervised Incomplete Multi-view Clustering. Data Science and Engineering. https://doi.org/10.1007/s41019-026-00348-8

Image Credits: AI Generated

DOI: 10.1007/s41019-026-00348-8

Keywords: incomplete multi-view clustering, contrastive learning, self-supervised learning, missing data imputation, unsupervised learning, machine learning, data mining, neighbor relationships, clustering algorithms, autoencoders, pseudo-labels, representation learning