Graph Feature Propagation for Incomplete Data Recovery

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2289 papers415 conceptsfrozen 2026-10-04

In many real-world machine learning scenarios, data is rarely "complete." A patient might miss a specific clinical test, a user might leave their profile attributes blank, or a camera might be occluded by an obstacle. Traditional approaches often handle these gaps using simple statistical fixes, like filling in the average value (mean imputation) or discarding the incomplete records entirely.

However, a growing body of research suggests that the relational structure of data—how entities connect to one another—contains enough redundant information to "repair" these missing pieces. By treating data points as nodes in a graph, researchers are using message-passing and diffusion mechanisms to propagate information from feature-rich regions to feature-sparse ones.

The Intuition: Topology as a Proxy for Features

The core insight behind this direction is structural homophily: the principle that connected entities tend to share similar characteristics. If we know the relationships between nodes (the graph topology), that structure can act as a functional proxy for the missing attributes.

As Xu Chen and colleagues demonstrate with the Node Attribute Neural Generator (NANG), it is possible to generate high-quality synthetic attributes even for nodes that have no observed features at all, simply by mapping their position in the graph to a latent attribute space. This works because a node's "neighborhood" provides a statistical context that constrains what its missing features are likely to be.

Feature Propagation as Diffusion

How does the information actually move? Many state-of-the-art methods frame this as a diffusion process. In On the Unreasonable Effectiveness of Feature propagation in Learning on Graphs with Missing Node Features, Emanuele Rossi and his team show that missing features can be recovered by minimizing Dirichlet energy.

Think of this like heat dispersing across a surface:

  1. Known features act as fixed heat sources (boundary conditions).
  2. Missing values are the cold spots.
  3. The algorithm iteratively "smooths" the features across the graph edges until the values of connected nodes are as similar as possible, while keeping the known values anchored.

This simple diffusion approach is surprisingly robust; Rossi found it could maintain high accuracy even when 99% of features were missing. Similar logic is applied in specialized fields like genomics, where Sukwon Yun uses scFP (single-cell Feature Propagation) to clean up noisy RNA sequencing data by diffusing known gene expression levels across cell-similarity graphs.

Bridging Modalities and Views

The problem of missing data becomes more complex in "multi-view" settings, where data comes from different sources (e.g., a person's image, their voice, and their text bio). If one "view" is missing, we face information asymmetry.

To solve this, researchers like Zhong Ji propose aligning the relationships between samples rather than the samples themselves. In MAP-Net, relationship data is transferred from a feature-rich modality to a feature-sparse one. If we know that Sample A and Sample B are similar in their "text" view, we can use that relationship to infer how they should look in their missing "image" view.

This "consensus" approach is a recurring theme:

  • Relationship Imputation: Instead of guessing raw pixel values, Shuping Zhao and Jie Wen argue in HLSCG_IMC that it is more effective to complete the affinity graph (the map of who is similar to whom) using information shared across available views.
  • Cross-Modality Transfer: In person re-identification, Yan Lu uses cm-SSFT to "borrow" traits from infrared images to fill in gaps in RGB images by looking at cross-modality neighbors.

Applications in High-Stakes Domains

This isn't just a theoretical exercise; it solves critical bottlenecks in data-starved fields:

  • Healthcare: Alberto García-Durán developed EP-MD, which uses message-passing to fill in missing Electronic Health Records. By looking at "clinically similar" peers, the model can predict patient outcomes even when specific lab results are missing.
  • Finance: For loan default prediction, Zihao Li uses MGCN to "fill in the blanks" of loan applications by pulling features from similar peers in a social-like graph.
  • Computer Vision: When a person is partially hidden, Guan'an Wang uses HOReID to "repair" occluded body parts by propagating information from visible regions through a graph of the human skeleton.

Summary of the Mechanism

The shift in this research area is from local imputation (looking only at the record itself) to relational imputation (looking at the record's place in the world). By leveraging the fact that data points do not exist in a vacuum, these models use the graph as a scaffold to reconstruct missing information, ensuring that the "hallucinated" data remains consistent with the global structure of the population.

Papers in this version

Showing 60 of 289 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 175 papers
  2. Version 1 · 2026-10-04 · 289 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_graph_based_feature_imputation_1ac38400,
  title = {graph-based feature imputation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-based-feature-imputation/rdr_1ac38400fb6a4dd68d1aee6fd8f6acb6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_graph_feature_propagation_for_incomplete_1ac38400_v1,
  title = {Graph Feature Propagation for Incomplete Data Recovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-feature-propagation-for-incomplete-data-recovery/rdr_1ac38400fb6a4dd68d1aee6fd8f6acb6/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}