Short Message Auditing Model Training With Active Sample Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing model training methods require large amounts of manually labeled data and struggle with the need for continuous iteration and optimization due to the addition of unlabeled samples that do not provide new information, leading to inefficiencies and high labeling costs.
Innovation Solution
A method involving sample reduction and active learning to select representative and informative samples for labeling, using strategies like maximum entropy and classifier voting to improve model performance, combined with techniques like cross-entropy and focal loss functions for training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a large amount of manually labeled data is used for model training, then model performance can be improved, but labeling costs and time consumption increase significantly
Solution Approach 1:
The system enables automatic model iteration by having the model select its own training samples from unlabeled data based on prediction probabilities and information value assessment, eliminating the need for manual sample selection and labeling prioritization
Solution Approach 2:
The patent changes the parameter of sample selection from random or manual selection to selection based on prediction probability and information value metrics, automatically identifying samples that will most improve model performance
2Adaptability or versatility
If unlabeled samples are continuously added to iterate and optimize the model, then model adaptability improves, but many samples do not provide new information leading to training inefficiency
Solution Approach 1:
The system implements a feedback mechanism where the model's prediction probabilities are used to assess the information value of unlabeled samples, creating a closed loop that guides sample selection for training to ensure each batch provides new information
Solution Approach 2:
Instead of using all available unlabeled samples, the system selectively processes only the subset of samples that meet the information value threshold, avoiding redundant training on samples that do not contribute new information
3Reliability
If manual labeling is performed continuously to supplement latest labeled data, then model can be iterated and optimized, but labeling costs increase
Solution Approach 1:
The model autonomously identifies which unlabeled samples are most valuable for training by evaluating prediction probabilities and information content, eliminating the need for manual assessment and prioritization of labeling tasks
Solution Approach 2:
The system performs preliminary assessment of unlabeled samples' information value before labeling, pre-filtering and ranking samples so that labeling resources are allocated only to high-value samples that will most improve model performance
Data Source
AI summary
The present application discloses a model training method, a short message auditing method and apparatuses, a device, and a storage medium, and relates to the field of artificial intelligence. A specific implementation solution of model training is: performing a sample reduction on first unlabeled samples to obtain second unlabeled samples; inputting the second unlabeled samples to a machine learning model for prediction, to obtain a probability corresponding to a result of predicting the second unlabeled samples; selecting a third unlabeled sample from the second unlabeled samples according to the probability; and training the machine learning model by using a third unlabeled sample after labeling. In embodiments of the present application, redundant samples are removed through a sample reduction, such that the selected samples have a certain degree of representativeness. In addition, a machine learning model is used to further select an informative sample with the most labeling significance for the current model by using an active learning technology, the cost of labeling is reduced.