Logic Rule Induction on Engineering Knowledge Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for learning first-order logic rules from data in engineering design domains face challenges due to a combinatorial search space, inefficiency, and inability to handle noise, especially in industrial applications where billions or trillions of candidate formulas exist, and current methods are limited to formulas with up to three predicates.
Innovation Solution
A system utilizing an agglomerative beam search method for disconnected knowledge graphs and graph neural networks for well-connected graphs to efficiently induce first-order logic rules, reducing the search space and handling noise without template constraints, enabling the derivation of longer logic formulas.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If brute-force search of all possible logic rule formulas is performed, then complete coverage of the search space is achieved, but computational intractability occurs due to billions or trillions of candidate formulas
Solution Approach 1:
The search space is segmented into disconnected knowledge graphs and well-connected knowledge graphs, with different search strategies applied to each type. This divides the massive search space into manageable portions, enabling complete coverage while maintaining computational efficiency.
Solution Approach 2:
The system dynamically adapts its search strategy based on the structure of the knowledge graph being analyzed. For disconnected graphs, it uses one search approach while using a different approach for well-connected graphs, allowing efficient navigation of the combinatorial space without exhaustive brute-force search.
2Productivity
If existing template-based methods are used to reduce search space, then computational feasibility is improved, but the methods become inefficient and unable to handle noise in real-world data
Solution Approach 1:
The patent replaces template-based mechanical search methods with a graph neural network approach that learns patterns from data. This substitution allows the system to handle noise in real-world data while maintaining computational feasibility, as the GNN processes the knowledge graph structure directly without relying on pre-defined templates.
3Reliability
If deep learning methods with chained logic variables and two-argument predicates are used, then noise handling capability is improved, but the methods are limited to formulas with up to three predicates which is inadequate for industrial applications
Solution Approach 1:
The patent extends the capability of deep learning methods from handling only three-predicate formulas to extracting logic formulas with any number of predicates. By changing the dimensional capacity of the model, it enables complex industrial applications while maintaining noise handling capabilities through the graph neural network approach.
4Loss of time
If the search space for logic formulas is reduced using templates, then processing time is decreased, but the search space reduction is insufficient for handling billions or trillions of candidate formulas
Solution Approach 1:
The search space is segmented by identifying disconnected knowledge graphs, which are then processed separately. This segmentation dramatically reduces the effective search space from billions or trillions of candidates to a manageable number, achieving both time reduction and comprehensive coverage.
Data Source
AI summary
System and method for logic rule formula induction on knowledge graphs for engineering system designs include receiving plurality of knowledge graphs for an engineering system. For a disconnected knowledge graph, agglomerative beam search is constrained to edges connected from node of interest, and candidate formulas are generated representing a respective edge found by the beam search engine, each formula constrained by a requirement of at least two arguments for defined formula chain length. Formula evaluation establishes whether each candidate formula is valid. Top ranked formulas are selected from the candidate formulas according to defined criteria. For well-connected graphs, a graph neural network is trained to predict first class for a query graph and second class for distractor graph. Counterfactual solver engine solves for minimum number of edits to query graph toward distractor graph which transforms predicted first class of the query graph to predicted second class.