Dual-Label Model Training with Inferred User Action Signals
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models often perform poorly due to the lack of extensive manually-labeled training data, especially in tasks where labeling is costly or privacy-constrained, and relying solely on noisy user actions for training can lead to models that fit noise rather than improving accuracy.
Innovation Solution
A machine learning model structure that combines explicit and inferred labels, using an encoder to map input data into a shared vector space, with separate classification layers for each type of label, and a training mechanism that prioritizes the order of selecting training examples to reduce noise impact.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manually-labeled training data is used to train machine learning models, then model accuracy is improved, but the cost and time required for data preparation increases significantly
Solution Approach 1:
The system performs preliminary actions by automatically generating inferred labels from user actions before the actual training process. This preliminary labeling using weak supervision signals (user interactions, clicks, views) prepares the training data in advance, reducing the need for time-consuming manual labeling while maintaining adequate training quality for machine learning models
Solution Approach 2:
An intermediary labeling system is introduced that uses weak supervision signals and heuristics to generate inferred labels. This intermediary layer bridges the gap between raw user actions and the structured labels needed for training, automatically creating training data without requiring direct human annotation for every example
2Quantity of substance
If noisy user actions are used for training, then the amount of training data increases, but the model may fit noise rather than improving accuracy
Solution Approach 1:
The system applies local quality by differentiating between different types of training data and applying different processing strategies. Explicitly-labeled examples are treated as high-quality data for initial training, while noisy inferred labels are treated as lower-quality data for subsequent fine-tuning. This localized quality assessment allows the model to learn from abundant data without being overwhelmed by noise
Solution Approach 2:
The system performs preliminary training on explicitly-labeled data before introducing noisy inferred labels. This preliminary action establishes a solid foundation of accurate knowledge first, then gradually incorporates noisier data in subsequent training iterations, allowing the model to distinguish signal from noise more effectively
3Measurement precision
If only explicitly-labeled examples are used for training, then label quality is high, but the amount of training data is insufficient
Solution Approach 1:
The system merges two distinct data sources: explicitly-labeled examples (high quality, low volume) and inferred labels from user actions (lower quality, high volume). By combining these complementary data sources in a multi-stage training process, the system achieves both sufficient data quantity and maintained label quality, overcoming the limitations of using either source alone
Data Source
Figure 1
Figure 2
Figure 3
AI summary
This document relates to training of machine learning models. One example method involves providing a machine learning model having a first classification layer, a second classification layer, and an encoder that feeds into the first classification layer and the second classification layer. The example method also involves obtaining first training examples having explicit labels and second training examples having inferred labels. The inferred labels are based at least on actions associated with the second training examples. The example method also involves training the machine learning model using the first training examples and the second training examples using a training objective that considers first training loss of the first classification layer for the explicit labels and second training loss of the second classification layer for the inferred labels. The method also involves outputting a trained machine learning model having the encoder and the first classification layer.