Hybrid Machine Learning and Expert System for Knowledge Base Completion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing knowledge base completion techniques for graph-structured knowledge bases either rely on symbolic AI or machine learning, struggling to effectively infer missing facts and provide recommendations in complex industrial AI applications where sufficient training data is limited, and require separate deployment and maintenance of expert systems and machine learning models.
Innovation Solution
A method and system that consolidates expert systems with machine learning through an active learning loop, using an expert system to verify candidate triples generated by a machine learning model, outputting novel facts for retraining, thereby improving generalization performance and reducing training time and deployment costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If machine learning techniques are used for knowledge base completion, then the system can learn from training data to solve reasoning tasks, but the approach requires sufficient training data which is often unavailable in complex industrial AI applications
Solution Approach 1:
The patent combines machine learning techniques with expert system techniques into a hybrid architecture. The machine learning model generates candidate facts while the expert system verifies them using domain knowledge rules, enabling the system to perform reasoning tasks without requiring large amounts of training data.
Solution Approach 2:
The expert system acts as an intermediary between the machine learning model and the final knowledge base completion results. It verifies candidate triples generated by the ML model using domain-specific rules, filtering out incorrect predictions and improving overall accuracy when training data is limited.
2Reliability
If expert system techniques are used for knowledge base completion, then the system can infer novel facts using logical rules, but the approach cannot effectively handle complex domains where hand-crafted logical rules are insufficient
Solution Approach 1:
The patent merges the strengths of expert systems (reliable logical reasoning) with machine learning (adaptability to complex patterns). The ML component handles complex domain patterns from available training data, while the expert system ensures reliable verification of inferred facts through logical rules.
Solution Approach 2:
The system implements a feedback mechanism where the expert system verifies candidate triples generated by the ML model. Correct verifications reinforce the ML model's learning, while rejections provide feedback for improvement, enabling the system to progressively enhance its performance in complex domains.
3Reliability
If separate expert systems and machine learning models are deployed for knowledge base completion, then each component can specialize in its strengths, but the deployment and maintenance effort increases
Solution Approach 1:
The patent integrates the expert system and machine learning model into a unified hybrid architecture where both components work together in a coordinated manner. This unified system maintains the specialized strengths of each approach while reducing overall deployment and maintenance complexity compared to separate systems.
Solution Approach 2:
The hybrid system serves multiple functions within a single architecture: the ML model performs pattern recognition and candidate generation, while the expert system provides verification and logical reasoning. This multi-functionality reduces the need for separate specialized systems.
4Reliability
If all possible candidate triples are verified by the expert system, then comprehensive novel facts can be inferred, but the computational time and resources are excessively high
Solution Approach 1:
Instead of verifying all possible candidate triples, the system applies partial verification by focusing on the most promising candidates generated by the ML model. The expert system verifies a selective subset of candidates, achieving sufficient completeness without the excessive computational cost of exhaustive verification.
Solution Approach 2:
The machine learning model serves itself by generating high-quality candidate triples that are more likely to be correct, reducing the verification burden on the expert system. This self-service capability allows the system to achieve comprehensive fact inference with reduced verification time.
Data Source
AI summary
An initially trained machine learning model is used by an active learning module to generate candidate triples, which are fed into an expert system for verification. As a result, the expert system outputs novel facts that are used for retraining the machine learning model. This approach consolidates expert systems with machine learning through iterations of an active learning loop, by bringing the two paradigms together, which is in general difficult because training of a neural network (machine learning) requires differentiable functions and rules (used by expert systems) tend not to be differentiable. The method and system provide a data augmentation strategy where the expert system acts as an oracle and outputs the novel facts, which provide labels for the candidate triples. The novel facts provide critical information from the oracle that is injected into the machine learning model at the retraining stage, thus allowing to increase its generalization performance.