Learning from Incomplete Multi-View Data

567 concepts330 papers1 proposal

In an ideal machine learning scenario, multi-view data—information gathered from different sensors, modalities, or feature sets—is perfectly aligned and complete. For instance, every video in a dataset would have a corresponding audio track, transcript, and metadata. In practice, however, sensors fail, users provide partial information, and data corruption occurs. This results in the missing view problem, where specific modalities are unavailable for a significant portion of the instances.

Standard multi-view algorithms typically assume that every data point is represented across all available perspectives. When this assumption is violated, these models cannot effectively aggregate information, leading to degraded performance. Research in this area focuses on how to leverage the "cross-view synergy" of available data to infer missing information or learn robust, unified representations despite these gaps.

Recovery through Matrix and Tensor Factorization

The most direct way to handle missing views is to treat the problem as a matrix or tensor completion task. If we view the entire dataset as a collection of matrices (one per view), the missing data creates "holes" in these structures.

Early approaches, such as OMVC by He et al., utilized weighted Nonnegative Matrix Factorization (NMF). By assigning zero weights to missing instances, the model learns a latent representation based only on observed data. To capture more complex, higher-order correlations between views, He et al. and Wen et al. proposed tensor-based frameworks. These methods treat the multi-view data as a 3D tensor and use CP decomposition or low-rank regularizations to "fill in the blanks," ensuring that the recovered data maintains manifold-space consistency across all views.

Dual Prediction and Contrastive Learning

Rather than just ignoring missing data, more recent frameworks treat representation learning and data recovery as "two sides of the same coin." The intuition is that if two views describe the same object, they must share a common underlying structure.

The COMPLETER framework, introduced by Lin et al., uses contrastive learning to maximize the mutual information between different views. By forcing the model to predict one view from another, it learns to recover missing features while simultaneously grouping related data points. Similarly, Wen et al. developed DMVG, which employs conditional diffusion models to generate missing views based on available ones. This generative approach allows the model to handle highly incomplete datasets by creating high-fidelity synthetic versions of the missing modalities.

Scalability via Anchor-Based Methods

A major challenge in incomplete multi-view clustering (IMVC) is computational complexity. Traditional methods often rely on kernel matrices or graph structures that grow quadratically with the number of samples ($n$). To address this, researchers have turned to anchor-based learning.

Instead of comparing every sample to every other sample, these methods learn a small set of representative "anchors." Wen and Zhu introduced a structure alignment module to ensure these anchors remain consistent across different views, even when data is missing. To further improve robustness, Yu et al. proposed DVSAI, which uses multi-scale anchors to capture complex data distributions. These techniques achieve linear computational complexity, allowing incomplete multi-view clustering to scale to datasets with hundreds of thousands of samples.

Handling Uncertainty and Weak Labels

In many real-world applications, missing views are not the only source of noise. Data might also have "weak labels" (missing tags) or unaligned samples.

Xie et al. proposed treating missing data not as a single "best guess" value, but as a distribution of possibilities. Their UIMC framework uses evidential deep learning to provide a "trustworthiness" metric for its predictions, allowing the model to be cautious when data is too sparse. Furthermore, Tan et al. introduced the iMVWL framework, which is the first to simultaneously tackle both missing views and weak labels by learning a shared discriminative subspace that captures local label correlations.

By shifting from simple imputation to joint learning frameworks that exploit cross-view consistency, these methods allow machine learning models to remain effective even when the available information is fragmented and noisy.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 348 papers
  2. Version 1 · 2026-10-04 · 347 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_incomplete_multi_view_data_69ea48eb,
  title = {incomplete multi-view data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/incomplete-multi-view-data/rdr_69ea48eb01bf445896c90976140d9118},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_incomplete_multi_view_data_69ea48eb_v2,
  title = {incomplete multi-view data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/incomplete-multi-view-data/rdr_69ea48eb01bf445896c90976140d9118/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake