Source-linked AI summary
Empowering Users in Graph Rule Mining via Large Language Models
Francesco Cambria, Francesco Invernici, Andrea Colombo, Anna Bernasconi
TL;DR
Graph association-rule mining with MINE GRAPH RULE requires expertise in graph theory and formal query languages, limiting access for domain experts. The paper uses prompted LLMs to generate, interpret, and rank graph rules. In 15-run tests, translations were manually validated at 13/15 for GPT and 15/15 for Gemini and Claude.
Problem
MINE GRAPH RULE and graph-based discovery require expertise in graph theory and formal query languages, creating an access barrier for domain experts.
Source: §1 · p. 2Method
Results
13 translations were correct with GPT and 15 with Gemini and Claude across 15 runs per model, based on manual validation.
Source: §2 · p. 2Takeaways & Limitations
The two-agent workflow can turn ranked rules into more refined ideas and subsequent mining of more specific, valuable rules.
Source: §4 · p. 4
Abstract
from arXiv · showhide
In the era of interconnected data, graphs have emerged as an effective abstraction for modeling complex systems in an intuitive format, especially with the rise of Property Graphs, which offer an intuitive and scalable way of navigating non-intuitive structures. In this context, graph mining techniques have been developed for testing complex graph-based rules, as the MINE GRAPH RULE operator, which, however, require users to have prior expertise both in graph theory and formal query language. In this work, we propose to bridge the gap between users and the graph-association rule-mining process by showing how Large Language Models (LLMs) can be easily prompted to formulate, refine, and interpret complex relational rules, directly producing MINE GRAPH RULE queries.
1 Introduction
The introduction frames property-graph mining as powerful but difficult for nontechnical users, then proposes LLMs as mediators for formulating, refining, and interpreting graph rules.
1 Introduction
- Property graphs represent entities and attribute-rich relationships in domains such as social networks, finance, and biology, but their complexity raises query-scalability and navigation challenges.The passage contrasts graph structures with tabular formats and emphasizes that connections carry important context.Source: §1 · p. 1
- Association rule mining finds associations in frequent topological patterns, with methods including GPAR, GARs, and PARM.Source: §1 · p. 1
- MINE GRAPH RULE provides formal syntax for extracting property-graph association rules and supports logic such as edge counts and multi-hop paths.It was implemented as a Neo4j stored procedure and enforces a strict structure for rule configurations.Source: §1 · p. 2
- The paper proposes LLMs to mediate graph-rule mining so domain experts can formulate, refine, and interpret rules without prior graph-theory or formal-query-language expertise.The proposal addresses the steep learning curve faced by researchers such as those in sociology or medicine.Source: §1 · p. 2
2 Mining a Graph database with LLMs
The section uses system-prompted LLMs to turn user interests and graph schemas into MINE GRAPH RULE queries, with preliminary manual validation across three models.
2 Mining a Graph database with LLMs
- The prompt frames the LLM as an association-rule mining and property-graph expert focused on Neo4j and Cypher, tasked with interpreting schema and user interests to form rules.It directs the model to choose anchors and itemset patterns.Source: §2 · p. 2
- The prompt includes operator examples for few-shot learning and uses a purchase schema linking Person and Item nodes through relationships classified by Category nodes.Figure 3 illustrates generated operators for varied user inputs.Source: §2 · p. 2
- 13 translations were correct with GPT and 15 with Gemini and Claude across 15 runs per model, based on manual validation against operator logic and parameters.Source: §2 · p. 2
3 Interpreting Graph Association Rules with LLMs
The section identifies the burden of reviewing numerous, hard-to-read graph rules and proposes a second LLM prompt to interpret and rank them for users.
3 Interpreting Graph Association Rules with LLMs
- Highly connected graph data can produce many rules for manual evaluation, while support and confidence may not identify the most meaningful results and raw rule outputs can be counterintuitive.The concern is especially relevant to researchers outside computer science and property-graph database communities.Source: §3 · pp. 2–3
- A secondary system prompt can translate rule outputs into accessible language and interpret confidence-1 or near-1 rules as functional graph constraints useful for repair and error detection.Source: §3 · p. 3
- The LLM can rank rules using logic tailored to users’ expertise and goals, enabling personalized filtering beyond simple statistical thresholds.The interpretation-and-ranking prompt is shown in Figure 4.Source: §3 · p. 3
4 Conclusion
The conclusion presents prompting as a way to bring graph association-rule mining closer to end users through two agents that support iterative refinement.
4 Conclusion
- The conclusion frames accessibility for nontechnical users as a central challenge because domain experts may struggle to interpret or apply patterns without graph-database expertise.Source: §4 · p. 3
- The two-agent workflow has one agent generate initial rules from general interests and another rank them using the user’s domain expertise.Source: §4 · p. 4
- Ranking can prompt more refined ideas or highlight interesting patterns, leading users back to the first agent to mine more specific and valuable rules.Source: §4 · p. 4