Downsampling Negative Signals in ML Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine-learned models trained using 'skip' interactions in job searches tend to be biased towards negative signals, leading to a decay where few or no job postings are presented to users, as the system assumes users are uninterested in most postings.
Innovation Solution
Implementing a skip logic with downsampling on negative signals in the training data set to limit the number of negatively labeled pieces of training data, reducing their impact on the model's bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If negative interaction signals (skips) are used extensively in training data, then the model learns user preferences more comprehensively, but the model becomes biased towards negative signals and decays to presenting few or no job postings
Solution Approach 1:
The patent changes the parameter of training data composition by dynamically adjusting the proportion of negative signals through downsampling. Instead of using all skip interactions, the system selectively subsamples negative examples to maintain a balanced training distribution, preventing model decay while preserving preference learning capability
Solution Approach 2:
The patent applies partial action by using only a subset of available negative interaction signals through downsampling. Rather than incorporating all skip interactions which would excessively bias the model, the system uses a controlled portion of negative signals to maintain model stability
2Quantity of substance
If all skip interactions are treated as negative signals in training, then more training data is available, but the majority of job postings are incorrectly treated as negative interactions
Solution Approach 1:
The patent changes the quantity parameter of training data by applying downsampling to negative signals. This maintains sufficient training data volume while adjusting the composition to reflect actual user intent more accurately, preventing mislabeling of potentially interesting job postings
3Ease of operation
If the model is trained to recognize user disinterest accurately, then fewer irrelevant job postings are shown, but the system assumes users are uninterested in most postings and presents very few results
Solution Approach 1:
The patent changes the parameter of training data balance by downsampling negative signals, which adjusts the model's sensitivity to negative interactions. This prevents over-filtering of job postings while maintaining the ability to identify truly irrelevant results, preserving both relevance and presentation volume
Data Source
AI summary
In an example embodiment, a skip logic using downsampling is applied to negative signals on a training data set fed to a machine-learning algorithm to train a machine-learned model. By downsampling the negatively labeled pieces of training data, the technical problem encountered in biasing the machine-learned model towards negative cases is overcome.