Embedding Updates With Feedback to Avoid Model Retraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional online concierge systems face challenges in maintaining accurate embeddings for users and items due to frequent updates, leading to impaired model performance and requiring computationally intensive retraining, which is time-consuming.
Innovation Solution
The system initializes updated embeddings by backpropagating error terms through existing embeddings, adjusts them based on gradients, and iteratively refines the adjustments to maintain model accuracy without full retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If embeddings are updated frequently to account for changes in item attributes or user characteristics, then the accuracy of the machine learning model is improved, but the feature space of the embeddings changes and impairs model performance
Solution Approach 1:
The system calculates a feedback term based on the change in model accuracy caused by the embedding update. This feedback term is then used to adjust the updated embedding, creating a closed-loop system that compensates for the disruption to the feature space while maintaining the benefits of frequent updates.
2Reliability
If the model is retrained when embeddings are updated to account for differences in feature spaces, then model performance is maintained, but the process is computationally intensive and time consuming
Solution Approach 1:
The system performs preliminary actions by calculating the feedback term and adjusting the embedding before the next model inference. This prevents the need for full retraining by proactively compensating for feature space changes, thereby maintaining model performance without the time cost of retraining.
Solution Approach 2:
Instead of retraining the entire model, the system changes the embedding parameters by adjusting them with the feedback term. This parameter adjustment approach maintains model performance while avoiding the computationally intensive process of full model retraining.
3Adaptability or versatility
If embeddings are updated to reflect recent changes in data, then the system adapts to new information, but the updated embeddings may have different feature spaces that degrade model accuracy
Solution Approach 1:
The system uses feedback from model accuracy measurements to guide the embedding update process. By calculating the feedback term based on accuracy changes and incorporating it into the adjusted embedding, the system maintains adaptability to new information while preserving feature space consistency and model accuracy.
Data Source
AI summary
An online concierge system uses a model to predict a user's interaction with an item, based on a user embedding for the user and an item embedding for the item. For the model to account for more recent interactions by users with items without retraining the model, the online concierge system generates updated item embeddings and updated user embeddings that account for the recent interactions by users with items. The online concierge system compares performance of the model using the updated item embeddings and the updated user embeddings relative to performance of the model using the existing item embeddings and user embeddings. If the performance of the model decreases, the online concierge system adjusts the updated user embeddings and the updated item embeddings based on the change in performance of the model. The adjusted updated user embeddings and adjusted updated item embeddings are stored for use by the model.