Domain-Adaptive Feature Interaction for Multi-Domain CTR Prediction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing CTR prediction models struggle to effectively handle multi-domain scenarios due to varying data distributions and resource-intensive separate modeling, leading to suboptimal performance and high maintenance costs.

Innovation Solution

A neural network model that determines domain features based on feature distribution, uses a dynamic network to convert these features, and employs a feature interaction network for accurate CTR prediction, considering both domain-specific and common characteristics.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If separate models are constructed for each domain, then prediction accuracy for domain-specific characteristics is improved, but device complexity and maintenance costs increase

Engineering Contradiction:
Improveprediction accuracyVSAvoidmodel complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent merges multiple domain-specific models into a single unified model that processes features from different domains simultaneously. The model integrates domain features, common features, and cross-domain interaction features in one architecture, eliminating the need to maintain separate models for each domain while preserving the ability to capture domain-specific characteristics through feature engineering and interaction terms.

Inventive Principle:
Principle #5Merging (Combining)

2Device complexity

If a unified model is used across multiple domains, then device complexity is reduced, but prediction accuracy deteriorates due to varying data distributions

Engineering Contradiction:
Improvemodel complexityVSAvoidprediction accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent applies local quality by introducing domain-specific feature processing within the unified model architecture. Different domains can have customized feature engineering, selective feature subsets, and domain-specific interaction terms that capture local characteristics. This allows the model to adapt to varying data distributions in each domain while maintaining a single unified structure, thereby preserving prediction accuracy without increasing overall model complexity.

Inventive Principle:
Principle #3Local quality

3Measurement precision

If domain-specific features are extensively processed, then prediction accuracy is improved, but computational resources and training time increase

Engineering Contradiction:
Improveprediction accuracyVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent segments the feature processing pipeline into distinct modules: domain feature extraction, common feature extraction, and cross-domain interaction feature generation. This segmentation allows for efficient parallel processing of different feature types and enables selective training strategies where computationally intensive interaction features can be pre-computed or approximated, reducing overall training time while maintaining the benefits of comprehensive feature processing for prediction accuracy.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4679294A1Data processing method and apparatus
Publication Date: 2026.01.14 HUAWEI TECH CO LTD
  • EP4679294A1 patent drawingFigure 1
  • EP4679294A1 patent drawingFigure 2
  • EP4679294A1 patent drawingFigure 3~4

AI summary

A data processing method and apparatus, a computer-readable storage medium, a chip, and a computer program product are provided. The method includes: determining a domain feature in a plurality of features based on distribution of each of the plurality of features of to-be-processed data; obtaining a plurality of corresponding converted representations by separately inputting a plurality of eigenvectors corresponding to the plurality of features to a dynamic network, where a dynamic network parameter of the dynamic network is obtained based on the domain feature; and obtaining a network output by inputting the plurality of converted features to a feature interaction network.