Graph Search Range Optimization for Learning Data Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for machine learning on graph structures face challenges in determining a suitable range of learning data from a huge knowledge base with limited resources, leading to suboptimal extraction of relevant nodes and edges, which affects learning accuracy.
Innovation Solution
A method that sets search ranges within a graph based on a set problem and node/edge relevance, using evaluation values from a learning model to optimize the extraction of learning data, incorporating processes like priority extraction, scaling, preferred terminal setting, and deletion delay to balance resource constraints and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all nodes or edges are set as learning targets, then learning accuracy may improve, but resource consumption (memory and processor) becomes excessive
Solution Approach 1:
The patent extracts only the relevant subset of nodes and edges from the huge knowledge base that are actually needed for solving the set problem. By using the learning model to evaluate relevance and the search method to identify important nodes/edges, the system takes out only the necessary portion of data rather than processing all data, thus reducing resource consumption while maintaining learning accuracy.
Solution Approach 2:
The patent applies local quality by treating different nodes and edges differently based on their relevance to the set problem. Instead of uniform processing, the system assigns different evaluation values and selection priorities to different parts of the graph based on their local characteristics and relationships to the problem, optimizing resource allocation to where it matters most.
2Quantity of substance
If the search range is limited by a small number of hops from the starting point, then resource consumption is reduced, but useful nodes located at distant positions cannot be extracted
Solution Approach 1:
The patent makes the search range dynamic rather than static. Instead of fixing the search depth to a predetermined number of hops, the system dynamically adjusts the search range based on the evaluation values generated by the learning model. The search continues as long as there are nodes and edges with sufficient relevance, allowing the search depth to adapt to the actual structure and relevance distribution in the knowledge base.
Solution Approach 2:
The patent incorporates feedback mechanisms where the learning model continuously evaluates nodes and edges during the search process. The evaluation values provide feedback that guides the search expansion, allowing the system to identify and pursue relevant nodes even at greater distances from the starting point while avoiding irrelevant branches, thus balancing search scope with resource constraints.
3Measurement precision
If the search range is expanded to include distant nodes, then learning accuracy improves, but resource consumption increases
Solution Approach 1:
The patent extracts only the relevant subset of nodes and edges from the huge knowledge base that are actually needed for solving the set problem. By using the learning model to evaluate relevance and the search method to identify important nodes/edges, the system takes out only the necessary portion of data rather than processing all data, thus reducing resource consumption while maintaining learning accuracy.
Solution Approach 2:
The patent applies local quality by treating different nodes and edges differently based on their relevance to the set problem. Instead of uniform processing, the system assigns different evaluation values and selection priorities to different parts of the graph based on their local characteristics and relationships to the problem, optimizing resource allocation to where it matters most.
Data Source
AI summary
A learning device sets each search range that is a part of a graph in which a plurality of nodes are connected by a plurality of edges, for each start node of one or more start nodes in the graph, on the basis of a set problem and relevance of an attribute of each of one or more nodes or edges connected to the each start node, generates an evaluation value of each of the one or more nodes and edges included in the each search range on the basis of a learning result by a learning model of the set problem, updates the one or more nodes and edges of the each search range on the basis of the evaluation value corresponding to the each search range and the relevance of the attribute corresponding to the each search range and the set problem, and performs a search.