Training Data Acquisition for Image Search Classifiers

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current search intent classifiers in online search systems rely heavily on manual labeling of training data, which is costly and ineffective in capturing the direct correlation between images and search queries, particularly for long-tail scenes, leading to suboptimal image classification and sorting.

Innovation Solution

A method and device for acquiring training data that involves selecting and filtering images based on user input, determining target-classification pairs, and grouping them to create high-quality training data, which includes filtering out untrusted clicks and evaluating accuracy to improve the classifier's performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual labeling is used to obtain training data for the classifier, then the classification accuracy can be improved, but the cost and time consumption increase significantly

Engineering Contradiction:

Improveclassification accuracyVSAvoidtime consumption

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The system enables users to automatically generate training data through their own search operations and image selections. Users perform searches, select relevant images from results, and the system automatically creates target-classification pairs without requiring manual labeling experts, thus making the training data generation process self-service oriented

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The system copies real user search behaviors and selections to create authentic training data. By replicating actual user interactions with search results and image selections, the system generates training samples that reflect genuine search intent patterns without requiring artificial manual creation

Inventive Principle:

Principle #26Copying

2Measurement precision

If manual labeling is used to obtain training data for the classifier, then the classification accuracy can be improved, but the cost increases significantly

Engineering Contradiction:

Improveclassification accuracyVSAvoidcost

Core Design Contradiction:

Measurement precisionVSLoss of energy

Solution Approach 1:

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The system uses inexpensive user interactions (clicks and selections) as disposable training data sources. Each user search and selection generates a training sample that can be immediately utilized, replacing the need for expensive manual labeling while maintaining data quality through automated processing

Inventive Principle:

Principle #27Cheap short-living objects (Disposable)

3Ease of manufacture

If traditional sorting methods are used based on text correlation and click numbers, then the sorting process is simple, but the direct correlation between image and search query is not captured effectively

Engineering Contradiction:

Improvesorting process simplicityVSAvoidimage-search correlation accuracy

Core Design Contradiction:

Ease of manufactureVSMeasurement precision

Solution Approach 1:

The system introduces user image selections as an intermediary element between the search query and the sorting process. Instead of directly correlating text with images, the system uses user selections of specific images from search results as a mediator to capture the true relevance relationship, which then feeds into the sorting mechanism

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The system replaces the mechanical text-matching approach with a behavior-based correlation method. Instead of relying on textual overlap and click metrics, the system substitutes user selection behaviors as the foundation for establishing image-query relationships, capturing semantic relevance that text correlation alone cannot detect

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

4Adaptability or versatility

If manual labeling is used to cover long-tail scenes, then the classification coverage can be improved, but the cost and time consumption increase significantly

Engineering Contradiction:

Improveclassification coverageVSAvoidtime consumption

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

The system enables continuous generation of training data for long-tail scenes through ongoing user search operations. As users continuously perform searches and select images across diverse topics, the system accumulates training samples for rare and long-tail categories without interruption, maintaining a steady flow of relevant training data

Inventive Principle:

Principle #20Continuity of useful action

Solution Approach 2:

Inventive Principle:

Principle #26Copying

Data Source

PatentUS11238310B2Training data acquisition method and device, server and storage medium

Publication Date: 2022.02.01 BAIDU ONLINE NETWORK TECH (BEIJIBG) CO LTD

AI summary

A training data acquisition method and device, a server and a storage medium are provided. The training data acquisition method is applied to a classifier and includes the following steps: obtaining an image search target according to an input of a user; providing images to the user according to the image search target, to display the images; and selecting at least one image from the displayed images, and determining a target-classification pair as training data according to the at least one image; where the target-classification pair includes the image search target and an entity-based classification of the at least one image.