Neural Network Entity Matching Without Retraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional machine learning models trained in a supervised learning paradigm are unable to match entities when new entities are added after training, as they assume a fixed set of labels, making them inefficient for dynamic environments where new labels are introduced.
Innovation Solution
A neural network-based machine learning model is trained using one-shot learning to generate a scalar confidence score for entity pairs, allowing it to match entities even if they were not accounted for in the original training data, by processing entity strings through an embedding layer, decomposable attention layer, and activation layer to determine the likelihood of a match.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning paradigm is used with fixed label sets, then training accuracy is improved, but adaptability to new entities deteriorates
Solution Approach 1:
The patent transforms the static classification problem into a dynamic similarity matching problem. Instead of mapping inputs to fixed labels, the system continuously computes similarity scores between query entities and database entities using neural networks. This dynamic approach allows the system to adapt to new entities without retraining, as the similarity computation is label-agnostic and works with any entity in the database.
Solution Approach 2:
The patent changes the fundamental parameter from discrete class labels to continuous similarity scores. By outputting a similarity score rather than a fixed class prediction, the system can handle new entities that were not present during training. The similarity score provides a graded measure of match quality that works regardless of whether the entity exists in the original training set.
2Device complexity
If traditional ML models are used, then model structure is simple, but ability to handle dynamic label sets deteriorates
Solution Approach 1:
The patent replaces traditional mechanical classification systems with neural network-based similarity computation. The neural network learns embeddings that capture semantic relationships between entities, enabling the system to compute similarity scores for new entities without requiring explicit classification rules or retraining. This substitution of neural representations for traditional classification mechanics enables dynamic adaptability.
3Measurement precision
If retraining is performed for each new entity, then matching accuracy is maintained, but computational efficiency deteriorates
Solution Approach 1:
The patent performs preliminary action by pre-training the neural network on available training data to learn entity representations and similarity computation. Once trained, the model can handle new entities immediately without requiring retraining. The preliminary training establishes the similarity computation mechanism that works universally for all entities, eliminating the need for repeated retraining operations.
Solution Approach 2:
The patent uses embedding layers to create vector representations (copies) of entities that capture their semantic meaning. These embeddings allow the system to compare new entities with existing ones through similarity computation without needing to retrain. The embedding copies enable efficient similarity matching that works for both training and test entities uniformly.
Data Source
AI summary
Methods, systems, and computer-readable storage media for a machine learning (ML) model and framework for training of the ML model to enable the ML model to correctly match entities even in instances where new entities are added after the ML model has been trained. More particularly, implementations of the present disclosure are directed to a ML model provided as a neural network that is trained to provide a scalar confidence score that indicates whether two entities in a pair of entities are considered a match, even if an entity in the set of entities was not accounted for in training of the ML model.