Text-Aware Contrastive Learning for Bridging Graph Components in a Joint Embedding Space

Read full text Summarise with VIRAT
Authors: Ahnaf Farhan, Moqsadur Rahman, Monika Akbar, Mahmud Shahriar Hossain
Cited by: 0
References: 53

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Abstract Graph-structured data across many domains consists of multiple heterogeneous components, each with its own features, that are related to one another even in the absence of explicit edges.In cybersecurity, for example, MITRE ATT\&CK and CWE provide complementary graph-structured knowledge of adversarial behaviors and system weaknesses, each with rich text-annotated nodes but no explicit links between the graphs. This disconnect prevents learning joint embeddings and limits downstream tasks such as similarity search, retrieval, and automated reasoning. We propose AttWeakBridge, a text-aware Graph Contrastive Learning (GCL) framework for embedding disconnected, text-rich graph components into a shared semantic space. Unlike conventional GCL methods that operate on augmented views of a single connected graph, AttWeakBridge performs contrastive learning across separate graph components, combining node text embeddings with structural signals via a dual-encoder GNN and an inter- / intra-graph triplet sampling strategy.Experimental results on real-world cybersecurity graphs (MITRE ATT\&CK and CWE) show that AttWeakBridge (i) improves similarity learning and cross-graph retrieval compared to text-only and structure-only baselines, and (ii) yields more coherent neighborhoods that align attack techniques with relevant weaknesses. This work contributes a general paradigm for embedding heterogeneous, disconnected graphs into a common semantic space.

Fields of Study & Subjects

Advanced Graph Neural Networks Graph Theory and Algorithms Topic Modeling

Publication details

Journal / source
Research Square
Publication date
26 March 2026
Language
EN
Document type
Preprint