[Paper Review] HyperJoin: LLM-augmented Hypergraph Link Prediction for Joinable Table Discovery

Shiyuan Liu, Jianwei Wang+4 more|arXiv (Cornell University)|Jan 3, 2026

Data Quality and Management0 citations

TL;DR

HyperJoin reframes joinable table discovery as link prediction on a hypergraph, using LLM-augmented inter-table edges and intra-table edges, with a Hierarchical Interaction Network and MST-based reranking to improve coherence and accuracy.

ABSTRACT

As a pivotal task in data lake management, joinable table discovery has attracted widespread interest. While existing language model-based methods achieve remarkable performance by combining offline column representation learning with online ranking, their design insufficiently accounts for the underlying structural interactions: (1) offline, they directly model tables into isolated or pairwise columns, thereby struggling to capture the rich inter-table and intra-table structural information; and (2) online, they rank candidate columns based solely on query-candidate similarity, ignoring the mutual interactions among the candidates, leading to incoherent result sets. To address these limitations, we propose HyperJoin, a large language model (LLM)-augmented Hypergraph framework for Joinable table discovery. Specifically, we first construct a hypergraph to model tables using both the intra-table hyperedges and the LLM-augmented inter-table hyperedges. Consequently, the task of joinable table discovery is formulated as link prediction on this constructed hypergraph. We then design HIN, a Hierarchical Interaction Network that learns expressive column representations through bidirectional message passing over columns and hyperedges. To strengthen coherence and internal consistency in the result columns, we cast online ranking as a coherence-aware top-k column selection problem. We then introduce a reranking module that leverages a maximum spanning tree algorithm to prune noisy connections and maximize coherence. Experiments demonstrate the superiority of HyperJoin, achieving average improvements of 21.4% (Precision@15) and 17.2% (Recall@15) over the best baseline.

Motivation & Objective

• Motivate joinable table discovery as a structure-aware problem in data lakes.
• Propose a hypergraph-based framework that captures intra-table and inter-table relations.
• Learn expressive column representations via a Hierarchical Interaction Network.
• Improve online ranking with a coherence-aware top-K selection and MST-based reranking.
• Demonstrate empirical gains over baselines on multiple benchmarks.

Proposed method

• Construct a hypergraph where columns are nodes with intra-table and LLM-augmented inter-table hyperedges.
• Encode initial column features from table name, column name, and cell values; fuse them into node embeddings.
• Apply a Hierarchical Interaction Network with Local Hyperedge Aggregation and Global Hyperedge Mixing for expressive representations.
• Use table-level and column-level positional encodings to address position agnosticism and global structural roles.
• Formulate online ranking as a coherence-aware top-K problem and solve with a greedy MST-based reranking to prune noisy connections.
• Demonstrate superiority through averaged improvements of 21.4% in Precision@15 and 17.2% in Recall@15 over state-of-the-art.

Experimental results

Research questions

• RQ1 How can joinable table discovery be improved by explicitly modeling intra-table and inter-table structure?
• RQ2 Can a hypergraph-based representation paired with LLM augmentation better propagate joinability signals across tables?
• RQ3 Does a coherence-aware top-K selection with MST reranking yield more coherent and accurate result sets than existing online ranking methods?

Key findings

• HyperJoin achieves average improvements of 21.4% in Precision@15 over the best baseline.
• HyperJoin achieves average improvements of 17.2% in Recall@15 over the best baseline.
• A hypergraph with intra-table and LLM-augmented inter-table hyperedges better captures structural context than isolated or pairwise representations.
• A Hierarchical Interaction Network enables effective global and local message passing on the hypergraph.
• A coherence-aware top-K ranking with an MST-based reranking module improves result coherence and reduces noisy connections.