🤖 machine learning

Joint Relational Database Generation via Graph-Conditional Diffusion Models

This paper introduces the Graph-Conditional Relational Diffusion Model (GRDM), a novel approach that leverages graph neural networks to jointly generate all tables in a relational database without imposing sequential order, thereby outperforming autoregressive baselines in capturing complex inter-table dependencies and achieving state-of-the-art fidelity.

Original authors: Mohamed Amine Ketata, David Lüdke, Leo Schwinn, Stephan Günnemann

Published 2026-05-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohamed Amine Ketata, David Lüdke, Leo Schwinn, Stephan Günnemann

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 Problem: The "Assembly Line" vs. The "Big Picture"

Imagine you are trying to recreate a massive, complex city from scratch. This city has different districts: a Residential District (people), a Commercial District (shops), and a Transport District (buses and trains). These districts are linked: people live in houses, shops are located on streets, and buses pick up people at stops.

The Old Way (Autoregressive Models):
Previous methods tried to build this city like a strict assembly line. They would say, "First, we must build all the houses. Once the houses are done, we can build the shops. Only after the shops are finished can we build the bus system."

This approach has three big flaws:

It's slow: You can't build the bus system until the houses are done. You can't work on everything at once.
It's rigid: If you need to fix a house later, you might have to tear down the shops and buses that were built based on the old house layout.
It misses the big picture: If a house is built wrong, the shops built next to it will also be wrong. The errors pile up, like a game of "telephone" where the message gets garbled by the time it reaches the end.

The New Solution: The "City Planner" (GRDM)

The authors propose a new method called GRDM (Graph-Conditional Relational Diffusion Model). Instead of an assembly line, they treat the whole city as a single, interconnected web (a graph) and build it all at once.

Here is how they do it, broken down into two main steps:

Step 1: Drawing the Blueprint (The Graph Structure)

Before building any actual buildings (data rows), the model first draws the "skeleton" of the city.

The Analogy: Imagine a master architect who knows exactly how many houses, shops, and buses usually exist and how they connect. They don't build the buildings yet; they just draw the map showing where the connections should be.
What the paper does: The model looks at the real database and learns the "degree distribution." This is a fancy way of saying: "On average, how many shops does one house connect to? How many buses stop at one station?" It then randomly generates a new map that follows these exact connection rules, ensuring the new city has the same structural shape as the real one.

Step 2: Filling in the Details (The Diffusion Model)

Once the map (the connections) is drawn, the model needs to fill in the details: the color of the houses, the names of the shops, the schedules of the buses.

The Analogy: Imagine the city is covered in thick fog (noise). The model starts with a blank, foggy map and slowly clears the fog, revealing the buildings one by one, but all at the same time.
How it works: This is where the "Diffusion" part comes in.
- In the real world, if you want to know what a specific shop sells, you look at the house next door and the bus stop nearby.
- The model does the same. To figure out the details of one "row" of data (a person), it looks at its immediate neighbors in the graph (the shops they visit, the buses they take).
- Because it looks at the neighbors, it understands the context. If the model sees a person connected to a "Luxury Car" shop, it knows that person probably has a high income. It doesn't have to guess in isolation; it uses the surrounding clues.

Why This is a Game Changer

1. No More "Assembly Line" Bottlenecks
Because the model looks at the whole graph at once, it can generate the houses, shops, and buses in parallel. It's like a team of painters working on every wall of a house simultaneously, rather than waiting for one wall to dry before painting the next.

2. Catching the "Long-Distance" Connections
In the old assembly line method, if a house in District A was connected to a bus in District B, which was connected to a shop in District C, the model often lost the connection between the House and the Shop.

The GRDM Advantage: Because the model "denoises" the data step-by-step, information travels through the network. Even if two things are far apart in the graph (like a house and a distant shop), the model eventually "hears" about each other through the chain of neighbors. It captures complex, multi-step relationships that previous models missed.

3. No "Order" Required
The old methods forced you to decide: "Do I build the houses first or the shops first?" The new method says, "It doesn't matter." It treats the database as a unified web, so you can generate any part of it without worrying about what came before it.

The Results: A Better Fake City

The authors tested this on six real-world databases (like customer records, movie ratings, and financial data). They compared their "City Planner" (GRDM) against the old "Assembly Line" methods.

The Verdict: The new method was significantly better at mimicking the real data, especially in how different tables (districts) related to each other.
The Proof: When they looked at complex connections (like "3-hop" relationships, where A connects to B, B to C, and C to D), the new model was much more accurate. It didn't just get the individual rows right; it got the relationships between them right.

Summary

Think of this paper as moving from building a city brick-by-brick in a strict line to using a smart, holistic blueprint that fills in the whole city simultaneously. By treating the database as a connected web and using a "fog-clearing" process to generate the data, the authors created a system that is faster, more flexible, and much better at understanding how different pieces of data rely on one another.

Problem
Relational databases (RDBs) organize data into multiple interlinked tables, storing the majority of the world's structured data. However, generating synthetic RDBs for privacy-preserving data release or data augmentation remains challenging. Existing approaches primarily rely on autoregressive factorizations, imposing a fixed table order (often based on the schema's topological order) and generating tables sequentially. This sequential generation introduces several limitations: it restricts parallelism, limits flexibility in downstream applications (such as missing data imputation), and forces simplifying conditional independence assumptions. These assumptions often lead to compounding errors and fail to capture complex, long-range dependencies (multi-hop correlations) between tables.

Methodology
The authors propose a fundamentally different, non-autoregressive approach called the Graph-Conditional Relational Diffusion Model (GRDM). Instead of generating tables sequentially, GRDM jointly models all tables in an RDB without imposing a specific table order.

Graph Representation: The RDB is represented as a directed, heterogeneous graph where nodes correspond to individual rows and edges encode primary-foreign key relationships. The graph structure $(V, E)$ captures the schema and connectivity, while node features $X$ represent the row attributes.
Two-Step Generation: The generation process is factorized into two steps:
- Structure Generation: A random graph generation algorithm creates the graph structure $(V, E)$ by preserving the empirical node degree distributions (specifically indegrees) of the real data. This step generates the primary and foreign key columns.
- Feature Generation: Given the fixed graph structure, the model generates the node features $X$ using a diffusion model. This is framed as modeling the conditional distribution $p(X|V, E)$ .
Graph-Conditional Diffusion: The core innovation is a diffusion model that conditions the denoising of each node on its local $K$ $K$ -hop neighborhood within the graph.
- Forward Process: Gaussian noise is added independently to all node features. Categorical variables are mapped to continuous space via label encoding to apply standard Gaussian diffusion.
- Reverse Process: A Graph Neural Network (specifically a heterogeneous Message-Passing GNN) predicts the noise added to a node based on the features of its $K$ -hop neighborhood. By treating the graph as undirected during message passing, nodes can interact with both parents and children.
- Long-Range Dependencies: Although the model conditions only on local neighborhoods ( $K=1$ in experiments), the iterative nature of the diffusion process allows information to propagate across the graph over multiple time steps, effectively capturing multi-hop correlations ($O(TK)$ hops) without explicit autoregressive conditioning.
Dimension Tables: Tables representing unique real-world entities (dimension tables) are treated as fixed metadata. Noise is not added to them; they remain fixed at $t=0$ , conditioning the generation of all other tables.

Key Contributions

Joint Modeling Framework: The paper introduces a framework that models the joint distribution of all tables in an RDB, eliminating the need for predefined table ordering or autoregressive factorization.
First Non-Autoregressive RDB Model: GRDM is presented as the first non-autoregressive generative model for RDBs, leveraging graph representations to jointly generate all row attributes.
Graph-Conditional Diffusion: The authors propose a novel diffusion mechanism that conditions denoising on graph neighborhoods, enabling the capture of complex inter-table dependencies.

Results
The authors evaluated GRDM on six real-world RDBs (Berka, Instacart 05, Movie Lens, CCS, California, and RelBench-F1) against state-of-the-art baselines, including SDV (statistical), ClavaDDPM (diffusion-based autoregressive), and two pipelines using single-table diffusion (SingleTable, Denorm).

Inter-Table Correlations: GRDM substantially outperformed autoregressive baselines in capturing multi-hop inter-table correlations. On the complex Berka dataset (up to 3-hop dependencies), GRDM improved 3-hop correlation metrics by 14.14% and 2-hop by 11.3% over the best baseline. On Instacart 05 and RelBench-F1, improvements on 2-hop correlations were significant (e.g., 473% relative improvement on Instacart 05).
Single-Table Fidelity: GRDM achieved state-of-the-art or comparable performance on single-table fidelity metrics (Cardinality, Column Shapes, Intra-Table Trends).
Ablation Studies:
- Replacing the sampled graph structure with the ground-truth graph yielded near-perfect structural fidelity, confirming the diffusion model's ability to learn attributes effectively.
- Converting GRDM into a sequential (autoregressive) variant resulted in significantly worse performance on both inter-table and single-table metrics, validating the benefit of joint modeling.
Scalability and Privacy: Experiments showed the model scales to larger database sizes without performance degradation. Privacy checks (Distance to Closest Record) indicated the model generates new records from the underlying distribution rather than memorizing training data.

Significance and Claims
The paper claims that GRDM addresses the fundamental limitations of sequential generation in RDB synthesis. By removing the constraint of table ordering and leveraging the relational structure through graph-conditional diffusion, the model achieves higher fidelity, particularly in capturing complex, long-range dependencies that autoregressive methods struggle with. The authors position this work as a foundation for scalable and expressive generative models for relational data, enabling more accurate synthetic data release and downstream analysis while preserving privacy. They acknowledge that while their graph generation algorithm is effective, more tailored approaches could further improve structural fidelity, and future work could explore extending GRDM to conditional generation tasks and differential privacy guarantees.