Abstract
We present an approach to data fusion that combines the interpretability of structured probabilistic graphical models with the flexibility of neural networks. The proposed method, lightweight data fusion (LDF), emphasizes posterior analysis over latent variables using two types of information: primary data, which are well-characterized but with limited availability, and auxiliary data, readily available but lacking a well-characterized statistical relationship to the latent quantity of interest. The lack of a forward model for the auxiliary data precludes the use of standard data fusion approaches, while the inability to acquire latent variable observations severely limits direct application of most supervised learning methods. LDF addresses these issues by utilizing neural networks as conjugate mappings of the auxiliary data: nonlinear transformations into sufficient statistics with respect to the latent variables. This facilitates efficient inference by preserving the conjugacy properties of the primary data and leads to compact representations of the latent variable posterior distributions. We demonstrate the LDF methodology on two challenging inference problems: (1) learning electrification rates in Rwanda from satellite imagery, high-level grid infrastructure, and other sources; and (2) inferring county-level homicide rates in the USA by integrating socio-economic data using a mixture model of multiple conjugate mappings.
Introduction
Data fusion is the process of integrating disparate data sources to infer unobserved latent quantities. In many real-world scenarios, researchers encounter two distinct data types. Primary data is typically high-quality but sparse, such as expensive ground surveys, for which a well-defined statistical model exists. In contrast, auxiliary data is ubiquitous but "messy"—think satellite imagery or socio-economic indicators. While auxiliary data clearly relates to the latent variable, defining a mathematical "forward model" (likelihood) for it is often intractable.
Standard Bayesian approaches struggle without this forward model, while supervised deep learning requires direct labels for latent variables that are often unavailable. Lightweight Data Fusion (LDF) bridges this gap by using neural networks to inform structured probabilistic models without requiring explicit generative models for the auxiliary data.
Methodology: Conjugate Mappings
The core innovation of LDF is the Conjugate Mapping. In Bayesian statistics, a conjugate prior allows for simple algebraic updates to a posterior. LDF leverages a neural network to transform auxiliary data $x$ into a set of "pseudo-observations" or sufficient statistics.
Because these mappings output parameters for exponential family distributions (e.g., the $\alpha$ and $\beta$ of a Beta distribution), the auxiliary data can be integrated into a Probabilistic Graphical Model (PGM) simply by adding its output to the statistics of the primary data. This "lightweight" approach allows for efficient, closed-form Bayesian inference.
Learning without Labels
A significant advantage of LDF is its training regime. The network is trained by maximizing the predictive likelihood of the primary data. Effectively, the neural network learns to extract features from the auxiliary data that best "explain" the observed primary data, using the primary data as a noisy supervisor. This allows the model to learn from auxiliary data even when direct ground-truth labels for the latent variables are missing.
Applications and Results
The authors evaluated LDF on two high-impact problems:
- Rwanda Electrification: Using geo-referenced distribution lines as primary data and nighttime lights/infrastructure as auxiliary data, LDF outperformed standard binomial and beta-binomial regression. It provided more certain posteriors for 60% of regions and a better fit for held-out survey data in 78.3% of cases.
- US Homicide Rates: Integrating FBI crime statistics with US Census socio-economic features, the authors employed a Mixture of Conjugate Mappings (LDF-MM). This approach captured multimodal behaviors that single-mapping models missed, achieving better held-out log-likelihood for 80.5% of US counties compared to gamma-Poisson baselines.
Discussion and Key Insights
LDF offers several advantages over traditional methods:
- Interpretability: Neural network outputs are directly interpretable as "pseudo-counts," quantifying the informational value of auxiliary data (e.g., how many virtual survey responses a satellite image represents).
- Uncertainty Quantification: Unlike point-estimate regression, LDF provides a full posterior distribution, enabling uncertainty-aware decision-making. This prevents the over-confidence often seen in Generalized Linear Models (GLMs).
- Efficiency: Inference relies on conjugate updates, making it highly parallelizable and computationally efficient post-training.
While LDF currently relies on the primary data following an exponential family distribution, it represents a significant step toward combining the strengths of deep learning and probabilistic modeling for complex data fusion tasks.
Summary updated