Data Subsampling for Recommendation Systems via Graph Conductance

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing data subsampling methods for recommendation systems face challenges due to model misspecification and inconsistencies between offline and online training, leading to inferior results and increased maintenance costs, especially in continuous integration and deployment scenarios.

Innovation Solution

The implementation of model-agnostic data subsampling methods that utilize the user-item graph to estimate interaction importance, smooth hardness scores, and combine with pilot model predictions to determine optimal subsampling rates, reducing dependencies across system modules and improving training efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If model-based data subsampling is used to reduce training data size, then productivity is improved, but reliability deteriorates due to model misspecification and offline-online inconsistencies

Engineering Contradiction:

Improvetraining efficiencyVSAvoidmodel performance

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent introduces an intermediary importance estimation mechanism that acts as a mediator between the pilot model and the target model. This intermediary uses a combination of pilot model predictions and graph-based hardness scores to determine subsampling rates, reducing the direct dependency on potentially misspecified pilot models and thereby improving reliability while maintaining productivity

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent implements a feedback mechanism where the importance estimation process continuously refines subsampling rates based on the relationship between pilot model predictions and actual target model performance. This feedback loop allows the system to adapt to model misspecification and offline-online inconsistencies, improving reliability without sacrificing training efficiency

Inventive Principle:

Principle #23Feedback

2Measurement precision

If pilot model dependencies are increased to improve subsampling accuracy, then measurement precision is improved, but device complexity increases due to additional system module dependencies

Engineering Contradiction:

Improvesubsampling accuracyVSAvoidsystem module dependencies

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies local quality by making the importance estimation process adaptive to local characteristics of different data regions. Instead of relying on a single global pilot model, the system uses local hardness scores derived from graph structures and pilot model predictions, allowing subsampling accuracy to be optimized locally without increasing overall system complexity

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent segments the importance estimation process into independent components: graph-based hardness scoring, pilot model prediction, and combined importance calculation. This segmentation allows each component to be developed and maintained independently, reducing system module dependencies while maintaining or improving subsampling accuracy

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS20230098656A1Data subsampling for recommendation systems

Publication Date: 2023.03.30 LEMON INC(GB)

AI summary

The present disclosure describes techniques for improving data subsampling for recommendation systems. A user-item graph associated with training data may be constructed. An importance of user-item interactions may be estimated via graph conductance based on the user-item graph. An importance of the training data may be measured via sample hardness using a pre-trained pilot model. A subsampling rate may be generated based on the importance estimated from the user-item graph and the importance measured by the pre-trained pilot model.