Reliable Multi-View Learning with Evidential Fusion

This direction is merged into another direction: evidential deep learning. Its last version stays readable and citable here.

Version 1 of 128 papers36 conceptsfrozen 2026-10-04

In traditional multimodal machine learning, we often assume that "more data is better." If we have an image, a text description, and a sensor reading, we typically concatenate their features or average their predicted probabilities. However, real-world data is rarely perfect. Sensors fail, images get blurred, and text can be ambiguous. In these cases, a standard model might confidently output a wrong answer because it lacks a mechanism to say, "I am seeing noise in this specific view."

A growing body of research, often termed Trusted Multi-View Classification (TMC), addresses this by treating each data source not as a simple feature vector, but as a source of evidence. By quantifying the uncertainty of each view before combining them, these models can dynamically prioritize reliable information and discount noise.

From Probabilities to Evidence

Standard neural networks use a Softmax layer to produce a probability distribution. While this tells us which class the model prefers, it doesn't distinguish between "I am certain this is a cat" and "I have no idea what this is, so I'm guessing 50/50."

To solve this, researchers like Zongbo Han and Changqing Zhang in Trusted Multi-View Classification propose modeling each view's output as a Dirichlet distribution. Instead of a single point estimate, the model outputs "evidence" parameters ($\alpha$) for each class.

The intuition is simple:

  • High Evidence: The model has seen enough patterns to support a specific class.
  • Low Evidence: The model is uncertain (vacuity of knowledge).

By parameterizing the output this way, the model can explicitly represent its own lack of knowledge. For example, in Multimodal Named Entity Recognition, Xiaoyang Zhong and colleagues use Normal-inverse Gamma distributions to allow the model to express "I am not sure" when image and text data conflict.

The Logic of Combination

Once each view has provided its "opinion" (belief masses) and its "uncertainty" (the remaining mass), the next challenge is fusion. How do you combine a "very certain" image with a "highly uncertain" text snippet?

The field relies heavily on Dempster-Shafer Theory (DST) and Subjective Logic. Unlike simple averaging, which would let a noisy view pull the final average toward the wrong class, DST uses a "Rule of Combination" to merge belief masses.

As Wei Liu and Thierry Denoeux demonstrate in Trusted Multi-View Deep Learning with Opinion Aggregation, this additive accumulation of evidence mathematically guarantees that as you add more informative views, the overall predictive uncertainty decreases. If a view is flagged as highly uncertain, the fusion rule naturally "shrinks" its influence, ensuring the final decision is dominated by the most trustworthy sources.

Handling Corruption and Incompleteness

This "evidence-first" approach is particularly powerful when data is missing or corrupted. In Exploring and Exploiting Uncertainty for Incomplete Multi-View Classification, Mengyao Xie and Changqing Zhang show that when a view is missing, the model can assign it a high uncertainty mass. During fusion, the system naturally ignores the "hallucinated" or missing data in favor of the observed, high-certainty views.

Similarly, in medical applications like arrhythmia classification, Mohd Ashhad uses evidential fusion to prioritize clean ECG signals over those corrupted by real-world noise. By converting class probabilities into evidence and conflict sets, the model can resolve disagreements between different views (like 1D time-series vs. 2D images) by looking at which one is more "certain" of its stance.

Why It Works: Sample-Adaptive Weighting

The core reason this direction succeeds where standard fusion fails is sample-adaptivity.

In a standard model, the "weight" given to the image view vs. the text view is usually a fixed parameter learned during training. If the image is usually better, the model learns to trust the image. But what if, for one specific sample, the image is blurry but the text is crystal clear? A fixed-weight model will still trust the blurry image.

Evidential models, such as the Event-Radar framework by Minnan Luo for fake news detection, calculate a "credibility coefficient" for every single input. This allows the model to:

  1. Quantify Reliability: Assess the quality of the specific input (e.g., is this image too blurry to be useful?).
  2. Discount Noise: Reduce the weight of that specific view for that specific sample.
  3. Provide a Safety Valve: If all views are uncertain, the model can output a high uncertainty score, signaling to a human that it cannot make a reliable decision.

By moving from "fusing features" to "fusing evidence," this research direction creates AI systems that are not just more accurate, but more robust to the messy, conflicting, and incomplete nature of real-world information.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_reliable_multi_view_learning_with_eviden_2c6cbefe,
  title = {Reliable Multi-View Learning with Evidential Fusion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reliable-multi-view-learning-with-evidential-fusion/rdr_2c6cbefeb3a94bf6ba045b06c18426ad},
  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_reliable_multi_view_learning_with_eviden_2c6cbefe_v1,
  title = {Reliable Multi-View Learning with Evidential Fusion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reliable-multi-view-learning-with-evidential-fusion/rdr_2c6cbefeb3a94bf6ba045b06c18426ad/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}