[Paper Review] Graph-based keyword search in heterogeneous data sources

Mhd Yamen Haddad, Angelos‐Christos G. Anadiotis+2 more|arXiv (Cornell University)|Sep 9, 2020

Data Management and Algorithms29 references4 citations

TL;DR

This paper proposes a graph-based keyword search algorithm for heterogeneous data sources, modeling diverse data types (relational, text, JSON, RDF) as a unified graph with bidirectional edges and sameAs links for entity matching. It enables users to discover cross-dataset connections without prior knowledge of schema or edge direction, preserving source provenance and supporting flexible scoring—offering the first solution for such integrated, heterogeneous keyword search with provenance preservation.

ABSTRACT

Data journalism is the field of investigative journalism which focuses on digital data by treating them as first-class citizens. Following the trends in human activity, which leaves strong digital traces, data journalism becomes increasingly important. However, as the number and the diversity of data sources increase, heterogeneous data models with different structure, or even no structure at all, need to be considered in query answering. Inspired by our collaboration with Le Monde, a leading French newspaper, we designed a novel query algorithm for exploiting such heterogeneous corpora through keyword search. We model our underlying data as graphs and, given a set of search terms, our algorithm nds links between them within and across the heterogeneous datasets included in the graph. We draw inspiration from prior work on keyword search in structured and unstructured data, which we extend with the data heterogeneity dimension, which makes the keyword search problem computationally harder. We implement our algorithm and we evaluate its performance using synthetic and real-world datasets.

Motivation & Objective

• To address the challenge of querying diverse, heterogeneous data sources (relational, text, JSON, RDF) in a unified manner for investigative journalism.
• To enable keyword search across multiple data models without requiring users to know schema or edge directions.
• To preserve source provenance by maintaining original node identities from each data source.
• To support flexible scoring without assuming specific score function properties, crucial for non-expert users.
• To design a scalable algorithm that finds meaningful connections between keywords across heterogeneous datasets.

Proposed method

• Model all heterogeneous data sources as a single integrated graph, preserving original node identities and source provenance.
• Use bidirectional edge traversal to allow queries to follow edges in either direction, regardless of original data model direction.
• Apply sameAs links to connect semantically equivalent entities across different data sources, with confidence scores between 0 and 1.
• Implement a grow-and-merge algorithm inspired by prior graph keyword search, adapted for bidirectional traversal and heterogeneous data.
• Support uncertain edges via confidence scores, enabling probabilistic reasoning in result construction.
• Avoid assumptions on score functions to maintain flexibility for real-world user needs in data journalism.

Experimental results

Research questions

• RQ1 How can keyword search be effectively performed across heterogeneous data sources with no shared schema or structure?
• RQ2 How can provenance be preserved when integrating data from multiple sources into a single queryable graph?
• RQ3 What algorithmic approach enables bidirectional traversal of edges without prior knowledge of data model direction?
• RQ4 How can uncertain entity matches (via sameAs links) be effectively used in constructing query answers?
• RQ5 How can the search algorithm remain scalable and efficient without assuming favorable score function properties?

Key findings

• The proposed algorithm successfully retrieves cross-dataset connections between keywords, even when edges are traversed in reverse direction relative to their original model.
• The method preserves source provenance by maintaining original node identities from each data source, enabling traceability of information origin.
• The approach supports uncertain links (via confidence scores) without requiring strong assumptions on the scoring function, enhancing usability for non-expert users.
• Evaluation on synthetic and real-world datasets shows the algorithm scales effectively despite the computational complexity of bidirectional, heterogeneous search.
• The solution outperforms prior work by enabling answers that span multiple data models—something not supported in existing keyword search systems.
• The system was validated in collaboration with Le Monde’s fact-checking team, demonstrating real-world applicability in data journalism.