Classifier training device, classifier training method, and classifier training program

The classifier learning device trains a binary classifier using positive and unlabeled data through density ratio estimation and class prior probability estimation, addressing the challenge of lacking negative examples and achieving high accuracy from a small dataset.

WO2025069146A1 Publication Date: 2025-04-03NIPPON TELEGRAPH & TELEPHONE CORP
0 Cites 0 Cited by

Patent Information

Application Number
PCT/JP2023/034757
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-09-25
Publication Date
2025-04-03

AI Technical Summary

Technical Problem

Conventional binary classifier training methods require large amounts of negative examples, which are difficult to collect, especially in real-world scenarios, making accurate training from a small amount of positive and unlabeled data challenging.

Method used

A classifier learning device that utilizes a related dataset to train a classifier generation model using positive and unlabeled data, estimating density ratios and class prior probabilities to generate a highly accurate classifier without relying on negative examples.

Benefits of technology

Enables highly accurate training of binary classifiers from a small amount of positive and unlabeled data, facilitating efficient and precise classification tasks.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

Provided is a classifier training device (1), wherein a training data input unit (11) acquires a related data set of a task related to a target task. A classifier generation model training unit (13) uses the acquired related data set to train a classifier generation model (14a) that generates a classifier for the task from P data and U data per task. A data input unit (21) acquires a target data set comprising P data and U data of the target task of the classifier. A training / classifying unit (23) uses the trained classifier generation model (14a) and the acquired target data set to generate a classifier for the target task.
Need to check novelty before this filing date? Find Prior Art