Surrogate Learning for Record Linkage Using Unlabeled Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high cost and time required for manually producing labeled data limits the practical application of machine-learning based classification systems, particularly in record linkage and paraphrase generation, as they rely heavily on large sets of labeled data for accuracy.
Innovation Solution
The development of semi-supervised learning techniques, specifically surrogate learning, which uses unlabeled data to supplement a small labeled sample by partitioning features into class-conditionally independent sets, allowing for the estimation of probabilities and decoupling the prediction problem into separate tasks, reducing the need for extensive labeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large sets of labeled data are used to train the classifier, then classification accuracy is improved, but time consumption and cost increase
Solution Approach 1:
The patent applies preliminary action by pre-processing unlabeled data through feature extraction and surrogate label assignment before the main classification task. The surrogate learning model is trained in advance on unlabeled data to generate pseudo-labels, which are then used to augment the limited labeled dataset, thereby preparing the data foundation ahead of time to reduce the need for extensive manual labeling.
Solution Approach 2:
The patent introduces an intermediary mechanism by using a surrogate learning model as a mediator between unlabeled data and the main classification task. This surrogate model generates intermediate surrogate labels that bridge the gap between unlabeled and labeled data, enabling the system to leverage large amounts of unlabeled data without requiring proportional manual annotation effort.
2Measurement precision
If large sets of labeled data are used to train the classifier, then classification accuracy is improved, but cost increases
Solution Approach 1:
The patent applies copying by creating surrogate labels that replicate the information content of true labels from unlabeled data. The surrogate learning model generates these copied labels through feature-based inference, allowing the system to multiply the effective size of the labeled dataset without incurring additional annotation costs, thereby maintaining classification accuracy while reducing the quantity of manually labeled data required.
3Ease of manufacture
If feature sets are partitioned into class-conditionally independent sets for surrogate learning, then the learning task is simplified and unlabeled data can be leveraged, but the complexity of the learning process increases
Solution Approach 1:
The patent applies segmentation by dividing the feature set into two conditionally independent subsets: features that are predictive of the target class and features that are not. This segmentation allows the surrogate learning model to focus only on the relevant feature subset when generating surrogate labels, simplifying the learning task by eliminating irrelevant features while maintaining the necessary complexity to capture meaningful patterns from unlabeled data.
Data Source
AI summary
A method of using unlabeled data to train a classifier is disclosed. In one embodiment related to record linkage, the method entails retrieving a set of candidate data records from a master database based on a least one update record. Next, a surrogate learning technique is used to identify one of the candidate data records as a match for the one update record. Lastly, the exemplary method links or merges the update record and the identified one of the candidate data records.