Multitable 5D Analysis — Relational Hypergraph Transformer (RHT)

Position & Design Paper Implementation Methodological and Conceptual Framework for 5D Multi-Table Analysis: A Unified Approach for Complex Data Reuse

Overview

This repository provides the reference implementation of the Relational Hypergraph Transformer (RHT) architecture described in:

Lansiaux E., Kazzi H., Loison A., Hammadi S., Chazard E. (2026). Methodological and Conceptual Framework for 5D Multi-Table Analysis: A Unified Approach for Complex Data Reuse.

The framework addresses five critical dimensions of complexity in multi-table data:

#	Dimension	Challenge
D1	Massive Volume	Scalability to millions of records across dozens of tables
D2	Multiplicity of Variables	Hundreds of columns per table
D3	High Categorical Cardinality	Thousands of distinct categories (e.g. ICD-10 codes)
D4	Multiple Tables & Relationships	Complex relational schemas (1:N, N:M, hierarchical)
D5	Repeated Temporal Measurements	Irregular time series across tables

Architecture

┌──────────────────────────────────────────────────────────┐
│                 Relational Hypergraph Transformer         │
│                                                          │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────────┐  │
│  │  Module 1    │  │  Module 2    │  │   Module 3      │  │
│  │  Hypergraph  │→ │  Temporal    │→ │   High-Card.    │  │
│  │  Construction│  │  Embeddings  │  │   Attention     │  │
│  └─────────────┘  └─────────────┘  └─────────────────┘  │
│         │                                     │          │
│         └──────────────┐  ┌───────────────────┘          │
│                        ▼  ▼                              │
│               ┌─────────────────┐                        │
│               │   Module 4      │                        │
│               │   Relational    │                        │
│               │   Discovery     │                        │
│               └─────────────────┘                        │
│                        │                                 │
│                        ▼                                 │
│               ┌─────────────────┐                        │
│               │  PentE Embedding│                        │
│               │  (5D unified)   │                        │
│               └─────────────────┘                        │
└──────────────────────────────────────────────────────────┘

Eight-Step Pipeline

1. Predictive Meta-Profiling — Automatic schema characterization
2. Automated Hypergraph Construction — Relational schema → hypergraph
3. Unified 5D Embedding (PentE) — Pentadimensional latent space
4. Relational Contrastive Learning — Contrastive training on relational structure
5. Dynamic Graph Rewiring — Task-adaptive relational graph reweighting
6. Relational Causal Inference — Cross-table causal discovery
7. Federated Multi-Table Learning — Privacy-preserving distributed training
8. Operationalization & Monitoring — Deployment, drift detection, explainability

Installation

git clone https://github.com/edlansiaux/multitable-5d-analysis.git
cd multitable-5d-analysis
pip install -e .

Requirements

• Python >= 3.10
• PyTorch >= 2.1
• DGL >= 1.1 (Deep Graph Library)
• NumPy, Pandas, Scikit-learn

Quick Start

from rht import RelationalHypergraphTransformer
from pipeline import MultiTablePipeline

# Load your relational tables
tables = {
    "patients": patients_df,
    "admissions": admissions_df,
    "diagnoses": diagnoses_df,
    "prescriptions": prescriptions_df,
    "lab_results": lab_results_df,
}

# Define schema relationships
relations = [
    ("patients", "patient_id", "admissions", "patient_id"),
    ("admissions", "hadm_id", "diagnoses", "hadm_id"),
    ("admissions", "hadm_id", "prescriptions", "hadm_id"),
    ("admissions", "hadm_id", "lab_results", "hadm_id"),
]

# Run the full 8-step pipeline
pipeline = MultiTablePipeline(tables, relations)
pipeline.run()

# Access PentE embeddings
embeddings = pipeline.get_pente_embeddings()

# Predict (e.g. hospital mortality)
predictions = pipeline.predict(task="mortality")

Repository Structure

multitable-5d-analysis/
├── rht/                          # Core RHT architecture
│   ├── __init__.py
│   ├── model.py                  # Main RelationalHypergraphTransformer
│   ├── modules/
│   │   ├── __init__.py
│   │   ├── hypergraph.py         # Module 1: Hypergraph Construction
│   │   ├── temporal.py           # Module 2: Multi-Scale Temporal Embeddings
│   │   ├── attention.py          # Module 3: High-Cardinality Attention
│   │   └── discovery.py          # Module 4: Differentiable Relational Discovery
│   ├── embeddings/
│   │   ├── __init__.py
│   │   ├── pente.py              # PentE: Pentadimensional Embedding
│   │   ├── semantic.py           # Semantic encoder
│   │   ├── categorical.py        # Hierarchical categorical encoder
│   │   └── volume.py             # Volume normalizer
│   ├── losses/
│   │   ├── __init__.py
│   │   ├── contrastive.py        # Relational-Temporal Contrastive Loss
│   │   └── relational.py         # Relational Discovery Loss
│   └── utils/
│       ├── __init__.py
│       ├── schema.py             # Schema detection utilities
│       └── temporal.py           # Temporal utilities
├── pipeline/
│   ├── __init__.py
│   ├── pipeline.py               # 8-step MultiTablePipeline
│   ├── profiling.py              # Step 1: Meta-Profiling
│   ├── rewiring.py               # Step 5: Dynamic Graph Rewiring
│   ├── causal.py                 # Step 6: Relational Causal Inference
│   └── federated.py              # Step 7: Federated Multi-Table Learning
├── evaluation/
│   ├── __init__.py
│   └── metrics.py                # 5D evaluation metrics (Section 7.2)
├── configs/
│   └── default.yaml              # Default hyperparameters
├── tests/
│   ├── __init__.py
│   └── test_modules.py           # Unit tests
├── examples/
│   └── mimic_iv_example.py       # MIMIC-IV walkthrough
├── setup.py
├── pyproject.toml
├── LICENSE
└── README.md

Benchmarks (MT-5D-Bench)

Planned benchmark datasets (Section 7.1):

Dataset	Tables	Records	Cardinality	Temporal
MIMIC-IV	26	15M+ measurements	10k+ ICD codes	Irregular vitals
Amazon Multi-Table	5	Millions	Hierarchical categories	Interaction TS
Financial Transactions	Multiple	High-frequency	Transaction codes	Millisecond

Benchmark suite will be hosted at: mt5d-benchmark.org

Anticipated Results (Design Targets)

Metric	SOTA (TGNN)	RHT (target)	Improvement
Rare Category F1	0.45	0.78	+73%
Relation Discovery Precision	0.67	0.92	+37%
Irregular TS Forecasting RMSE	1.23	0.89	-28%
Training Time (100 tables)	48h	12h	-75%
Memory Usage	64 GB	24 GB	-62%
Overall 5D Score	0.58	0.83	+43%

Warning: These are research targets, not confirmed empirical results. See Section 7.4 of the paper for the validation plan.

Citation

@article{lansiaux2026multitable5d,
  title={Methodological and Conceptual Framework for 5D Multi-Table Analysis:
         A Unified Approach for Complex Data Reuse},
  author={Lansiaux, Edouard and Kazzi, Hugo and Loison, Aur{\'e}lien
          and Hammadi, Slim and Chazard, Emmanuel},
  year={2026},
  institution={Lille University Hospital, Lille Centrale Institute}
}

License

MIT — see LICENSE.

Authors

• Edouard Lansiaux — Emergency Department, Lille University Hospital
• Hugo Kazzi — Lille Centrale Institute
• Aurelien Loison — Lille Centrale Institute
• Pr. Slim Hammadi — Lille Centrale Institute, CRISTAL UMR CNRS 9189
• Pr. Emmanuel Chazard — Department of Public Health, EA 2694, ULR 2694-METRICS