[Paper Review] Compiling Relational Database Schemata into Probabilistic Graphical Models

Sameer Singh, Thore Graepel|arXiv (Cornell University)|Dec 5, 2012

Data Management and Algorithms13 references3 citations

TL;DR

This paper proposes a method to automatically compile relational database schemata into fully-Bayesian probabilistic graphical models by modeling each table as a mixture model and using foreign key relationships to link component indicators across tables. The approach enables scalable, accurate inference for missing value prediction, outlier detection, and clustering, outperforming traditional join-based models on synthetic and real-world data with linear scalability.

ABSTRACT

Instead of requiring a domain expert to specify the probabilistic dependencies of the data, in this work we present an approach that uses the relational DB schema to automatically construct a Bayesian graphical model for a database. This resulting model contains customized distributions for columns, latent variables that cluster the data, and factors that reflect and represent the foreign key links. Experiments demonstrate the accuracy of the model and the scalability of inference on synthetic and real-world data.

Motivation & Objective

• To automate the construction of probabilistic graphical models from existing relational database schemata without requiring machine learning expertise.
• To leverage domain experts' schema design knowledge—such as foreign key relationships and attribute types—for building accurate, generative models.
• To enable scalable inference for missing value prediction, outlier detection, and relational queries using variational message passing.
• To demonstrate that schema-based models outperform traditional join-based single-table models in accuracy, especially with missing data.
• To evaluate the model’s clustering quality and scalability on real-world datasets like MovieLens and Xbox TrueSkill.

Proposed method

• Each table is modeled as a finite mixture model, with latent component indicators for each row and data-specific distributions (Gaussian, Bernoulli, or discrete) per attribute.
• Foreign key relationships are modeled by making the component indicator of a referencing table depend on the component indicator of the referenced row via a gate mechanism.
• Foreign key uncertainty is modeled as a discrete distribution, enabling prediction of missing or incorrect foreign key links.
• The full model is constructed iteratively by processing tables in topological order of the foreign key dependency graph, starting from leaf tables.
• Variational message passing in Infer.NET is used for efficient inference, with linear complexity in the number of rows when foreign keys are observed.
• The model supports probabilistic relational queries and marginal posterior inference over missing values using the compiled graphical model.

Experimental results

Research questions

• RQ1 Can a relational database schema be automatically compiled into a fully-Bayesian, generative probabilistic graphical model with minimal human intervention?
• RQ2 Does modeling foreign key relationships through component indicator dependencies improve accuracy in missing value prediction compared to standard join-based approaches?
• RQ3 Can the resulting model produce meaningful, interpretable clustering of rows without additional domain-specific assumptions?
• RQ4 Is the inference process scalable to large real-world databases with tens of thousands of rows?
• RQ5 How does the model perform in terms of robustness when a significant proportion of data values are missing?

Key findings

• The schema-based model achieved significantly lower RMSE on missing value prediction than a single-table model formed by joining all tables, especially when up to 50% of values were missing.
• On the MovieLens 100k dataset, inference runtime scaled linearly with the number of ratings, confirming the model’s efficiency and practical scalability.
• The model successfully clustered Xbox players into three distinct skill levels—bad, good, and excellent—based solely on match outcomes and foreign key links, demonstrating meaningful latent structure discovery.
• The approach maintained high accuracy even with high levels of missing data, showing robustness to data quality issues.
• The use of variational message passing enabled efficient inference, avoiding the intractability of Gibbs sampling in models with deterministic gates and strong dependencies.
• The model’s ability to predict missing foreign keys was effective, as foreign key uncertainty was explicitly modeled through discrete distributions over potential references.