Search Model Training with Knowledge Graph Negative Samples
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image search methods lack effective fine-granularity classification and rely solely on positive sample data sets for training, leading to unstable model performance and inaccurate predictions, making them unsuitable for high-precision searches.
Innovation Solution
A method involving the construction of a knowledge graph to determine additional query information for creating negative sample data sets, which are used in conjunction with positive sample data sets for training the search model, employing inversion pair learning to enhance convergence speed and prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional image search methods rely solely on positive sample data sets for training, then the training process is simple, but the model performance becomes unstable and prediction accuracy decreases
Solution Approach 1:
The patent applies inversion by constructing negative sample data sets through inverting the traditional approach: instead of only collecting positive samples that match query intent, the system uses knowledge graphs to identify and construct negative samples (images that do not match query intent) by traversing knowledge graph relationships from query keywords. This inverted sampling strategy resolves the contradiction by improving model stability through diverse training data while managing complexity through systematic knowledge graph-based construction.
Solution Approach 2:
The patent applies preliminary action by pre-construcing knowledge graphs and pre-processing query keywords before training. The system extracts keywords from queries, traverses the knowledge graph to identify related concepts and construct negative samples in advance, and prepares both positive and negative sample data sets before model training begins. This preliminary preparation resolves the contradiction by ensuring stable model performance through comprehensive data preparation while organizing complexity into manageable preprocessing steps.
2Measurement precision
If conventional image search methods use only positive sample data sets, then data collection is easier, but fine-granularity classification precision is insufficient
Solution Approach 1:
The patent applies segmentation by dividing the training data set construction into distinct segments: positive sample collection and negative sample construction. The system segments the knowledge graph traversal process into keyword extraction, concept identification, and negative sample generation stages. This segmentation resolves the contradiction by enabling precise fine-granularity classification through dedicated negative sample construction while organizing data collection into manageable segments rather than requiring overwhelming undifferentiated data quantities.
Solution Approach 2:
The patent applies local quality by creating specialized negative sample data sets tailored to specific query types and knowledge graph domains. Instead of using generic diverse data, the system constructs negative samples with specific local characteristics relevant to each query context by traversing knowledge graph relationships locally from query keywords. This resolves the contradiction by providing the precise local quality needed for fine-granularity classification while avoiding the need for overwhelming global data diversity.
3Measurement precision
If additional query information is determined using knowledge base traversal, then prediction accuracy improves, but processing time increases
Solution Approach 1:
The patent applies preliminary action by pre-traversing the knowledge graph and pre-construcing negative sample data sets before model training. The system extracts keywords from queries, traverses the knowledge graph to identify related concepts and construct negative samples in advance, and prepares comprehensive training data before the actual model training begins. This resolves the contradiction by achieving high prediction accuracy through thorough knowledge graph traversal while avoiding time loss during training by completing data preparation beforehand.
Solution Approach 2:
The patent applies beforehand cushioning by preparing extensive negative sample data sets in advance to cushion against the time-consuming nature of knowledge graph traversal. The system constructs a buffer of preprocessed negative samples that can be used during model training without requiring real-time knowledge graph queries, thus cushioning the training process from time delays while maintaining high prediction accuracy through the comprehensive negative sample coverage.
Data Source
AI summary
A method and apparatus for training a search model, and a method and apparatus for searching for a target object are provided. The method may comprise: obtaining, based on sample query information, a first sample data set used to train the search model. The method may further comprise: determining, based on a pre-constructed knowledge base, additional query information associated with the sample query information. In addition, the method may further comprise: obtaining, based on the additional query information, a second sample data set used to train the search model, wherein the second sample data set is different from the first sample data set. Moreover, the method may further comprise: training the search model based on the sample query information, the first sample data set, and the second sample data set.