Knowledge Graph Explanation via Neighborhood Sampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Knowledge graph embedding models struggle to predict links associated with new, unknown entities, requiring significant time and resources for retraining and failing to provide effective explanations for predicted links to end users.
Innovation Solution
A prediction platform that generates candidate responses, scores them based on a knowledge graph, determines neighborhoods with minimal quality loss, and creates reasoning graphs to explain candidate responses, using techniques like embedding approximation and neighborhood sampling to handle unknown entities efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If knowledge graph embedding models are retrained to predict links for new unknown entities, then prediction accuracy is improved, but training time and computing resources are significantly consumed
Solution Approach 1:
The system performs preliminary actions by pre-computing and storing embeddings for known entities before new entities arrive. When a new entity needs to be predicted, the system uses these pre-computed embeddings along with neighborhood sampling to quickly generate predictions without full retraining, thus avoiding the time penalty while maintaining accuracy.
Solution Approach 2:
The system segments the knowledge graph into neighborhoods around known entities. Instead of processing the entire graph during prediction, it only processes relevant local neighborhoods, dramatically reducing the computational burden and time required for predictions involving new entities while preserving prediction accuracy.
2Measurement precision
If knowledge graph embedding models are retrained to predict links for new unknown entities, then prediction accuracy is improved, but computing resources are significantly consumed
Solution Approach 1:
The system segments the knowledge graph into localized neighborhoods rather than processing the entire graph. This segmentation reduces the computational scope to only relevant portions, dramatically lowering the computing resources and energy required for predictions while maintaining prediction accuracy through focused local analysis.
Solution Approach 2:
The system applies partial action by using neighborhood sampling that processes only a subset of relevant entities rather than the complete knowledge graph. This partial processing approach reduces computing resource consumption while still achieving accurate predictions through representative sampling of the neighborhood.
3Loss of information
If explanations for predicted links are generated to improve interpretability, then user understanding is improved, but system complexity increases
Solution Approach 1:
The system extracts and highlights only the most relevant reasoning paths and entities from the neighborhood that contribute to the prediction. By taking out and presenting only the essential explanatory elements rather than the complete reasoning process, the system improves interpretability while avoiding the complexity of presenting all possible reasoning pathways.
Solution Approach 2:
The system applies local quality by providing detailed explanations only for the specific local neighborhood relevant to the prediction, rather than providing uniform comprehensive explanations for all predictions. This localized explanation approach improves user understanding of specific predictions without increasing overall system complexity.
Data Source
AI summary
A device receives a knowledge graph and an ontology for the knowledge graph, and receives a query for information associated with the knowledge graph. The device generates candidate responses to the query, and assigns scores to the candidate responses based on the knowledge graph. The device identifies a particular candidate response, of the candidate responses, based on the scores for the candidate responses, and determines, based on the knowledge graph, a neighborhood of the particular candidate response. The device generates knowledge graph embeddings for the neighborhood of the particular candidate response, and determines a particular neighborhood, with a smallest loss of quality, based on the knowledge graph embeddings. The device generates a reasoning graph based on the ontology and the particular neighborhood, and generates an explanation of the particular candidate response based on the reasoning graph. The device performs an action based the explanation of the particular candidate response.