Transfer Learning for Machine Learning Model Training Bias
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional online concierge systems face biases in training machine learning models due to how items are displayed, affecting user interactions, leading to decreased accuracy and user engagement.
Innovation Solution
The system generates item and user embeddings, trains a model using labeled training data, and incorporates exploration training data by randomly displaying items to users, modifying specific model parameters to offset display biases while conserving computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If items are displayed to users based on predicted likelihood of interaction, then user engagement is improved, but training data accuracy deteriorates due to display biases
Solution Approach 1:
The patent segments training data into two distinct sources: exploitation data (from biased display based on predictions) and exploration data (from random display). This segmentation allows the system to maintain separate training streams that address different needs - exploitation data maintains user engagement while exploration data ensures training accuracy by eliminating display biases
Solution Approach 2:
The patent changes the display parameter from prediction-based ordering to random ordering for exploration data collection. This parameter change eliminates the bias introduced by predictive display algorithms, allowing the model to learn from interactions that are not influenced by position or popularity biases
2Measurement precision
If alternative training data with random item display is used to offset biases, then training data accuracy is improved, but user engagement deteriorates due to increased user inputs required
Solution Approach 1:
The patent applies partial action by using random display only for a portion of training data collection (exploration data) rather than for all user interactions. The system maintains prediction-based display for regular operations to preserve user engagement, while selectively using random display only when collecting exploration training data
3Measurement precision
If the model is trained with exploration training data to offset display biases, then prediction accuracy is improved, but computational resources increase
Solution Approach 1:
The patent segments the model training into two phases: initial training with exploitation data and subsequent fine-tuning with exploration data. This segmentation allows the system to use computational resources efficiently by applying the computationally intensive exploration data training only after the initial model is already trained
Solution Approach 2:
The patent implements dynamic training where the system adaptively switches between using exploitation data and exploration data based on model performance and resource availability. The training process is made flexible to adjust the proportion of exploration data used, balancing accuracy improvement with computational cost
Data Source
AI summary
An online system uses a trained model predicting likelihoods of a user performing a specific interaction with items to order or to rank items for display to the user. The online system trains the model using interactions by users with items displayed by the online system. However, selection, popularity, and position from display of the items affects the model during training. To improve the model, the online system further trains the model using additional training data obtained from displaying items to users in different orders. The further training is done on a limited portion of the model, such as a limited number of layers of the model, to improve the model performance while reducing an amount of additional data to acquire to further train the model.