Knowledge Graph Negative Example Filtering for Training Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The training efficiency for embedding in knowledge graphs is low, resulting in poor prediction accuracy due to the arbitrary generation of negative example graph datasets, which includes unlinked data pieces that are not relevant for training.
Innovation Solution
The information processing apparatus generates negative example graph datasets by selecting only objects linked to the same property as in positive example graphs, excluding those that are not relevant, thereby improving the training process by focusing on actual unlinked data for training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If negative example graph datasets are generated arbitrarily including unlinked data pieces, then the training dataset coverage is increased, but the training efficiency decreases and prediction accuracy deteriorates
Solution Approach 1:
The patent extracts only the relevant portion of negative examples - specifically, unlinked objects that share the same property/predicate as positive examples. This selective extraction removes irrelevant unlinked data pieces while retaining useful training signals, thereby improving training efficiency without sacrificing dataset coverage.
Solution Approach 2:
The patent applies local quality by differentiating between relevant and irrelevant negative examples based on their relationship to the property being learned. Instead of treating all unlinked data equally, it focuses training on locally relevant cases where the object is unlinked but shares the same predicate, creating higher quality training signals in critical areas.
2Quantity of substance
If negative example graph datasets include unlinked data pieces, then more training samples are available, but the prediction accuracy becomes poor due to irrelevant training data
Solution Approach 1:
The patent extracts only the relevant subset of negative examples by filtering for unlinked objects that share the same property/predicate as positive examples. This extraction removes irrelevant unlinked data that would dilute the training signal, thereby maintaining prediction accuracy while still providing sufficient training samples.
Solution Approach 2:
The patent changes the selection parameter for negative examples from arbitrary selection to selection based on predicate matching. By changing how negative examples are parameterized and selected, the training data quality improves, leading to better prediction accuracy.
3Quantity of substance
If all unlinked data pieces are used for training, then the training data volume increases, but the training process becomes less effective due to inclusion of irrelevant data
Solution Approach 1:
The patent extracts only the relevant negative examples by taking out unlinked objects that share the same property/predicate as positive examples. This selective extraction maintains adequate training data volume while removing irrelevant unlinked data, thereby preserving training effectiveness.
Solution Approach 2:
The patent applies partial action by including only the necessary portion of negative examples - those unlinked objects with matching predicates - rather than all possible unlinked data. This partial inclusion is sufficient for effective training without the harm of excessive irrelevant data.
Data Source
AI summary
A storage medium storing an information processing program that causes a computer to execute a process that includes acquiring a graph dataset that includes graphs each having a subject, a predicate, and an object from a knowledge graph; generating a negative example graph dataset that includes one or more negative example graphs, a predicate of each negative example graph being same as a predicate of positive example graphs, an object of the negative example graph being different from an object of the positive example graphs, the negative example graph being excluded from the negative example graph dataset when the object of the negative example graph is different from each object to which a predicate of any other of the positive example graphs is linked; and training for embedding in the knowledge graph by using the positive example graph dataset and the negative example graph dataset.