📊 statistics

Bayesian Joint Additive Factor Models for Multiview Learning

This paper introduces two Bayesian joint additive factor models, JFR and JAFAR, which utilize novel cumulative shrinkage process priors to effectively integrate multiview data, infer complex dependencies, and improve outcome prediction while maintaining interpretability and uncertainty quantification.

Original authors: Niccolo Anceschi, Federico Ferrari, David B. Dunson, Himel Mallick

Published 2026-01-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Niccolo Anceschi, Federico Ferrari, David B. Dunson, Himel Mallick

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Multi-View" Problem

Imagine you are trying to understand a complex mystery, like why a specific car engine is making a strange noise. You have three different mechanics looking at it:

Mechanic A listens to the engine sounds (Audio).
Mechanic B looks at the oil and fuel chemistry (Chemistry).
Mechanic C inspects the electrical wiring (Electronics).

This is what the paper calls Multiview Learning. In medicine, this is like looking at a patient's DNA, their proteins, and their metabolites all at once. The goal is to combine these different "views" to predict an outcome (like when labor will start) and understand how the different parts of the body are connected.

However, there are three big problems with this approach:

Too much noise: Some data is just static; it doesn't tell you anything useful.
Too much data, too few people: You might have thousands of measurements (features) but only a few dozen patients (samples).
Confusing signals: It's hard to tell if a signal comes from a problem shared by all three mechanics (a shared engine issue) or just a glitch in one specific mechanic's tools (a view-specific issue).

The Solution: Two New Statistical Models

The authors propose two new "detective tools" (statistical models) to solve this. Think of them as two different ways to organize the clues.

1. JFR (Joint Factor Regression): The "Group Hug" Approach

Imagine all three mechanics huddle together and agree on a single list of "suspects" (latent factors) that explain the noise.

How it works: It assumes there is one master list of causes that affects all three views simultaneously.
The Catch: It treats everything as a big, messy pile. It's good at finding the general pattern, but it's hard to tell which mechanic contributed which specific clue. It's like a group hug where everyone is mixed up together.

2. JAFAR (Joint Additive Factor Regression): The "Specialized Team" Approach

This is the paper's main innovation. Instead of one big huddle, JAFAR splits the work into two teams:

The Shared Team: These are the "suspects" that affect all three views (e.g., a real engine problem).
The Specialized Team: These are "suspects" that only affect one view (e.g., a glitch in the audio equipment that has nothing to do with the engine).
Why it's better: It keeps the shared clues separate from the specific noise. This makes the results much easier to understand (interpretable) and faster to calculate.

The Secret Sauce: The "CUSP" Prior

How does the model know which clues are important and which are just noise? The authors invented a special rule called CUSP (Cumulative Shrinkage Process).

The Analogy: Imagine a row of light switches.
- Standard methods might flip switches randomly, leaving too many lights on (too many factors).
- CUSP is like a smart switchboard that says: "If the first few switches are dim, the next ones must be even dimmer, and eventually, they must be OFF."
- It automatically turns off the unnecessary factors, shrinking the noise down to zero.

For the JAFAR model, they created a smarter version called D-CUSP (Dependent CUSP).

The Problem: Without this, the model might get confused. It might think a "Specialized" clue is actually "Shared," or vice versa. It's like the mechanics arguing over who owns a specific tool.
The Fix: D-CUSP forces the rules to be strict. It says, "A factor is only 'Shared' if at least two mechanics agree on it. If only one mechanic sees it, it stays in the 'Specialized' pile." This ensures the model doesn't get confused about what is shared and what is unique.

Testing the Tools

The authors tested these tools in two ways:

The Simulation Lab (Fake Data): They created fake datasets where they knew the "truth" (they knew exactly which factors were shared and which were specific).
- Result: Their new tools (JFR and JAFAR) correctly identified the shared and specific factors. Other existing tools got confused, mixing them up or creating too many fake factors. JAFAR was also much faster to run.
The Real World Test (Labor Prediction): They applied the model to a real medical dataset involving 53 women. They had three types of data:
- Immunome (immune cells)
- Metabolome (metabolites)
- Proteome (proteins)
- Goal: Predict how many days until labor starts.
- Result: The new models predicted the timing of labor better than the competitors. They also managed to handle the "too much data, too few people" problem better by using a "tempered" version of their tool (JAFAR-T), which acted like a filter to reduce the number of factors to a manageable size without losing accuracy.

The Bottom Line

The paper introduces a new way to combine different types of data (like DNA, proteins, and metabolites) to make better predictions.

JFR is a solid, all-in-one approach.
JAFAR is the star of the show: it separates the "team effort" signals from the "individual noise," making the results clearer, faster to compute, and easier to interpret.
They also fixed a common math problem where models get confused about what is shared and what isn't, using their new D-CUSP rule.

The result is a toolkit that helps doctors and scientists see the forest and the trees when looking at complex, multi-layered biological data.

Technical Summary: Bayesian Joint Additive Factor Models for Multiview Learning

Problem Statement
The paper addresses the challenge of analyzing multiview data, where multiple types of high-dimensional, noisy, and heterogeneous measurements (e.g., genomics, metabolomics, proteomics) are collected on the same set of samples. A primary motivation is precision medicine, where integrating these views is crucial for understanding patient heterogeneity and predicting clinical outcomes. However, standard approaches face significant hurdles:

Signal-to-Noise Variability: Different views often possess vastly different signal-to-noise ratios, which standard fusion methods (early or late) fail to account for.
Interpretability and Identifiability: Existing methods struggle to disentangle variation that is shared across views from variation that is specific to a single view, particularly in high-dimensional, low-sample-size ( $p \gg n$ ) settings.
Computational and Statistical Limitations: Current Bayesian formulations often impose overly rigid constraints, introduce bias, or fail to simultaneously estimate the number of factors while reliably separating shared and view-specific components. Furthermore, additive factor models suffer from specific non-identifiability issues where view-specific factors can be confounded with shared factors.

Methodology
The authors propose two complementary Bayesian factor regression models:

Joint Factor Regression (JFR): A baseline model that captures combined variation across all views using a single set of latent factors ( $\eta_i$ ). It employs Independent Cumulative Shrinkage Process (I-CUSP) priors on the loading matrices. This allows the model to adaptively shrink unnecessary factors to zero, learning the effective rank of the shared structure without pre-specifying it.
Joint Additive Factor Regression (JAFAR): A more nuanced model that explicitly decomposes variation into two components:
- Shared Factors ( $\eta_i$ ): Fully or partially shared across at least two views.
- View-Specific Factors ( $\phi_{mi}$ ): Unique to a specific modality $m$ .
  To address the non-identifiability inherent in additive decompositions (where a view-specific factor could be mathematically re-expressed as a sparse shared factor), the authors introduce a Dependent Cumulative Shrinkage Process (D-CUSP) prior. This prior enforces a constraint during the Gibbs sampling adaptation step: a factor is included in the shared component if and only if it is active in at least two views. This ensures a measure-zero probability of misallocating view-specific factors to the shared set.

Key Technical Contributions

Novel Priors (I-CUSP and D-CUSP): The paper extends the CUSP prior (Legramanti et al., 2020) to the multiview setting. The D-CUSP specifically addresses the identifiability of additive models by inducing dependence across views to prevent the "misallocation" of factors, a problem not solved by standard orthogonality constraints in a Bayesian framework.
Tempered Extensions (JFRT and JAFART): Recognizing that standard CUSP priors may infer too many factors in extreme $p \gg n$ scenarios (where the inferred rank exceeds the sample size $n$ ), the authors propose tempered versions. These utilize fractional posteriors to treat the sample size $n$ as the effective sample size for rank selection, rather than the feature dimension $p$ .
Efficient Inference: The authors develop Gibbs samplers that exploit the structured sparsity of the JAFAR model. By performing joint updates for latent factors and loadings, they achieve computational costs that scale more favorably than non-partitioned representations, particularly for large $p$ .
Post-Processing: A modified Multiview MatchAlign procedure is introduced to resolve rotational, label, and sign ambiguities. This adapts the Varimax rotation to respect the composite structure of shared and specific components, preventing dominant views from distorting cross-view sparsity patterns.

Results
The methods were evaluated through simulation studies and a real-world application:

Simulation (Unsupervised): In separating shared and view-specific components, JAFAR with D-CUSP successfully recovered the true factor structure. Competing approaches (NAIVE and FULL-D priors) failed to correctly separate components, leading to inflated shared ranks and poor covariance reconstruction. JFR and JAFAR outperformed competitors (BIP, BSFP) in reconstructing both intra-view and inter-view correlation matrices.
Simulation (Supervised): JFR and JAFAR achieved competitive or superior predictive accuracy (MSE and $R^2$ ) compared to state-of-the-art competitors (Cooperative Learning, IntegratedLearner, BIP, BSFP). Crucially, this accuracy was coupled with accurate reconstruction of the underlying dependence structures. JAFAR demonstrated a significant computational advantage, requiring roughly 55% less runtime than JFR while maintaining similar predictive performance.
Real Data Application (Labor Onset Prediction): The models were applied to predict time-to-labor onset using immunome, metabolome, and proteome data ( $n=53$ $n = 53$ , $p \approx 6000$ $p \approx 6000$ ).
- Standard JFR and JAFAR inferred ranks exceeding the sample size, a known issue in extreme high-dimensional settings.
- The tempered versions (JFRT, JAFART) successfully reduced the inferred number of factors to levels comparable to BSFP while retaining the adaptive Bayesian framework.
- In terms of prediction, JAFAR and its tempered variant outperformed BSFP, Cooperative Learning, and IntegratedLearner in out-of-sample tests.
- JAFART provided better-calibrated 90% credible intervals (reduced overcoverage) compared to the untempered versions, attributed to reduced estimation noise from fewer latent factors.

Significance and Claims
The paper claims that JFR and JAFAR provide a fully Bayesian framework capable of simultaneously estimating the number of factors, separating shared from view-specific variation, and providing accurate uncertainty quantification. The authors emphasize that:

JAFAR is the preferred choice for high-dimensional settings due to its computational efficiency and enhanced interpretability, as it isolates view-specific signals (whether genuine or artifacts) from shared biological signals.
The D-CUSP prior is essential for the identifiability of additive models, solving a specific non-identifiability problem that plagues other additive factorization approaches.
The tempered formulations offer a practical solution for extreme $p \gg n$ regimes where standard factor models might infer spurious high-rank structures.
The open-source R package jafar implements these methods, facilitating their application in precision medicine and other multiview learning contexts.

The authors remain modest, noting that while JFR and JAFAR are essentially equivalent in terms of latent representation accuracy, JAFAR's structural decomposition offers distinct advantages for inference and scalability. They also acknowledge that the tempered versions are optional adaptations for specific extreme scenarios rather than a complete replacement for the original formulations.