Interpretable multi-source data fusion through Latent Variable Gaussian Process

Wei Chen, Northwestern University, Sandipp Krishnan Ravi, Yigitcan Comlek, Arjun Pathak, Vipul Gupta, Rajnikant Umretiya, Andrew Hoffman, Ghanshyam Pilania, Piyush Pandita, Sayan Ghosh, Nathaniel Mckeever, Wei Chen, Liping Wang

TL;DR

This paper introduces a "source-aware" data fusion framework that uses Latent Variable Gaussian Processes (LVGP) to combine data from multiple, often conflicting, sources like research papers, patents, and in-house experiments. By mapping categorical data sources into a physically interpretable latent space, the model not only improves prediction accuracy in sparse-data scenarios but also provides a way to quantify how much different data sources disagree with one another.

What Problem Does This Paper Address?

In fields like materials science, researchers rarely have the luxury of a single, massive, high-quality dataset. Instead, they must "fuse" data from various sources: a few in-house experiments, some numbers from a 1990s patent, and a handful of data points from a recent journal article.

The challenge is that these sources are rarely equal. They might use different experimental setups, have varying measurement errors, or involve "hidden" physical parameters that aren't reported in the text. Standard machine learning models usually treat all data points as if they came from the same distribution (source-unaware), which leads to "noisy" models with high uncertainty and poor predictive power. The authors aim to create a model that can ingest all this data while remaining "aware" of where each point came from and how much to trust it.

Methods

The authors leverage the Latent Variable Gaussian Process (LVGP). While standard Gaussian Processes (GPs) are excellent for modeling continuous data with uncertainty, they struggle with categorical variables (like "Source A" vs. "Source B") because there is no natural "distance" between categories.

The LVGP approach works by: 1. Latent Mapping: It treats each data source as a categorical variable and implicitly maps it into a low-dimensional (usually 2D) continuous "latent space." 2. Statistical Inference: During training, the model learns the coordinates of these sources in the latent space. If two sources are placed close together, it means the model found their underlying "mapping behavior" (the relationship between inputs and outputs) to be similar. 3. Dissimilarity Metric ($D$): The authors introduce a normalized metric based on the Euclidean distance in this latent space. This allows researchers to pick a "Ground Source" (like their own trusted lab data) and mathematically rank how similar other external sources are to it. 4. Source-Specific Prediction: Because the model knows the "coordinates" of each source, it can generate smoother, source-specific response surfaces, effectively filtering out the noise introduced by lower-quality data.

Datasets

The framework was validated using four distinct case studies:

  • Mathematical Parabola: A synthetic problem using four different equations (one "Ground" and three "Perturbed") to test if the model could recover the known relationships.
  • 2D Ackley Function: A standard optimization benchmark used to demonstrate how adding external data—even if slightly different—can actually reduce uncertainty on the primary source of interest.
  • FeCrAl Alloys: A real-world materials dataset (380 points) focused on "thermal aging embrittlement." It combined in-house GE Research data with seven external literature sources.
  • SmCoFe Alloys: A dataset (71 points) focused on the magnetic energy product of alloys, fusing data from two patents and one peer-reviewed paper.

Key Insights and Ideas

  • The Latent Space is a "Physics Proxy": The most striking finding is that the latent space isn't just a mathematical trick; it often reflects underlying physical reality. In the FeCrAl study, sources that used similar experimental procedures clustered together, while "outliers" in the latent space corresponded to sources with known differences in manufacturing.
  • More Data Isn't Always Better: The authors show that some data sources can actually hurt model performance if they are too dissimilar from the target domain.
  • Targeted Source Selection: Using their dissimilarity metric, the authors demonstrated that by removing the most "distant" sources in the latent space, they could improve prediction accuracy on their target data by up to 20%. This provides a principled way to perform "data cleaning" at the source level.
  • Confidence through Fusion: Even when external data is slightly "off," the LVGP can use the correlations between sources to build a more confident model of the ground truth than a model trained on the ground truth alone.

Key Results

  • Accuracy Boost: In the parabola problem, the LVGP achieved a Testing NRMSE (Normalized Root Mean Squared Error) of $11.66 \times 10^{-4}$, compared to $0.207$ for a standard source-unaware GP—a massive improvement in precision.
  • FeCrAl Case Study: The "Source-Aware" LVGP reduced error by roughly 25% compared to a standard GP. When the authors used the latent space to prune "bad" sources (the LVGP-Target model), they saw an additional 7% improvement over a model trained only on high-quality in-house data.
  • Uncertainty Reduction: In the SmCoFe magnet study, the LVGP's predicted uncertainty was three times lower than a single-source model, proving that "fusing" even small amounts of patent data can significantly stabilize a model.

Limitations and Future Work

  • Scalability: The authors note that the current LVGP approach struggles when the number of sources exceeds 15. As the number of sources grows, the number of latent variables to estimate becomes too high for the model to converge easily.
  • Heterogeneous Inputs: Currently, the model assumes all sources share the same input parameters (e.g., all sources must report Chromium and Aluminum content). Future work aims to extend this to "heterogeneous" cases where different sources might report different sets of variables.
  • Large Datasets: For very large datasets, the authors suggest moving toward Sparse-Variational LVGP to handle the computational load of Gaussian Process regression.

Summary updated

BibTeX
@article{university2025interpretable,
  author = {Wei Chen, Northwestern University and Sandipp Krishnan Ravi and Yigitcan Comlek and Arjun Pathak and Vipul Gupta and Rajnikant Umretiya and Andrew Hoffman and Ghanshyam Pilania and Piyush Pandita and Sayan Ghosh and Nathaniel Mckeever and Wei Chen and Liping Wang},
  title = {{Interpretable multi-source data fusion through Latent Variable Gaussian Process}},
  journal = {Engineering Applications of AI},
  year = {2025},
  doi = {10.1016/j.engappai.2025.110033},
}
Figure