Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SiMa: Effective and Efficient Matching Across Data Silos Using Graph Neural Networks

Christos Koutras, Rihan Hai, Kyriakos Psarakis, Marios Fragkoulis, Asterios Katsifodimos

TL;DR

SiMa addresses cross-silo column matching by learning from intra-silo column relationships via Graph Neural Networks. It casts inter-silo matching as a link-prediction task on per-silo relatedness graphs, using GraphSAGE embeddings and an MLP predictor, with negative sampling and incremental training to boost effectiveness and efficiency. The approach achieves higher accuracy than state-of-the-art schema matching and column-representation baselines while requiring significantly fewer resources, demonstrating practical applicability for federated data environments. Real-world experiments on NYC and LA OpenData benchmarks show substantial performance gains and scalable computation, enabling more effective data integration across silos.

Abstract

How can we leverage existing column relationships within silos, to predict similar ones across silos? Can we do this efficiently and effectively? Existing matching approaches do not exploit prior knowledge, relying on prohibitively expensive similarity computations. In this paper we present the first technique for matching columns across data silos, called SiMa, which leverages Graph Neural Networks (GNNs) to learn from existing column relationships within data silos, and dataset-specific profiles. The main novelty of SiMa is its ability to be trained incrementally on column relationships within each silo individually, without requiring the consolidation of all datasets in a single place. Our experiments show that SiMa is more effective than the - otherwise inapplicable to the setting of silos - state-of-the-art matching methods, while requiring orders of magnitude less computational resources. Moreover, we demonstrate that SiMa considerably outperforms other state-of-the-art column representation learning methods.

SiMa: Effective and Efficient Matching Across Data Silos Using Graph Neural Networks

TL;DR

SiMa addresses cross-silo column matching by learning from intra-silo column relationships via Graph Neural Networks. It casts inter-silo matching as a link-prediction task on per-silo relatedness graphs, using GraphSAGE embeddings and an MLP predictor, with negative sampling and incremental training to boost effectiveness and efficiency. The approach achieves higher accuracy than state-of-the-art schema matching and column-representation baselines while requiring significantly fewer resources, demonstrating practical applicability for federated data environments. Real-world experiments on NYC and LA OpenData benchmarks show substantial performance gains and scalable computation, enabling more effective data integration across silos.

Abstract

How can we leverage existing column relationships within silos, to predict similar ones across silos? Can we do this efficiently and effectively? Existing matching approaches do not exploit prior knowledge, relying on prohibitively expensive similarity computations. In this paper we present the first technique for matching columns across data silos, called SiMa, which leverages Graph Neural Networks (GNNs) to learn from existing column relationships within data silos, and dataset-specific profiles. The main novelty of SiMa is its ability to be trained incrementally on column relationships within each silo individually, without requiring the consolidation of all datasets in a single place. Our experiments show that SiMa is more effective than the - otherwise inapplicable to the setting of silos - state-of-the-art matching methods, while requiring orders of magnitude less computational resources. Moreover, we demonstrate that SiMa considerably outperforms other state-of-the-art column representation learning methods.
Paper Structure (20 sections, 4 equations, 8 figures, 4 tables, 2 algorithms)

This paper contains 20 sections, 4 equations, 8 figures, 4 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Approach Overview
  3. Related Work
  4. Schema Matching
  5. Dataset Discovery & Semantics Types
  6. Problem Definition
  7. GNNs for Matching Data Silos
  8. Preliminary: GNNs
  9. Modeling Data Silos as Graphs
  10. Profiles as Initial Features
  11. Training GNNs for Matching Silos
  12. SiMa's Pipeline
  13. Optimization Techniques
  14. Negative Sampling Strategies
  15. Incremental Training
  16. ...and 5 more sections

Figures (8)

  • Figure 1: Three typical data silos in the banking industry.
  • Figure 2: SiMa overview: (a) depicts data silos and their column matches which are transformed into relatedness graphs ((b)-\ref{['sec:model']}), where nodes represent columns and receive their initial features from a tabular data profiler (\ref{['sec:profile']}). Then, negative edges are being sampled from each relatedness graph as shown in (c) (\ref{['sec:sampling']}) and a link prediction model is being trained based on an incremental training scheme depicted in (d) (\ref{['sec:incre_train']}). Finally, using the trained model we are able to predict relationships among columns from different silos as depicted in (e).
  • Figure 3: Relatedness graph of the Insurance data silo.
  • Figure 4: Strategies for negative edge sampling on the relatedness graph of the insurance data silo.
  • Figure 5: NYC OpenData
  • ...and 3 more figures

Theorems & Definitions (4)

  • Definition : Data silos
  • Definition : Intra-relatedness and Inter-relatedness
  • Definition : Relatedness graph
  • Definition : Positive edges and negative edges