Deep Learning Revenue-Per-Click Prediction With Meta-Feature Integration

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional performance prediction for web advertisements does not adequately incorporate non-structural information such as item-type, price, hierarchy, and image features, limiting the accuracy and comprehensiveness of predictions.

Innovation Solution

A deep learning-based framework that utilizes meta features like item-type information, item price, item hierarchy, and item images, combined with historical traffic features, to generate performance metrics through a multilayer neural network, enabling real-time adjustments using an online learning service.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional performance prediction based on historical traffic features is used, then the prediction system is simple to implement, but the prediction accuracy is limited due to not incorporating non-structural information

Engineering Contradiction:

Improveprediction accuracyVSAvoidsystem complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the feature extraction process into multiple specialized components: a meta-feature extractor for non-structural information (item-type, price, hierarchy, images), a traffic feature extractor for historical data, and a multilayer perceptron for integration. This segmentation allows each component to specialize in specific data types while maintaining overall system manageability despite increased complexity.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent adds a new dimension to the prediction system by incorporating non-structural meta-features (item-type, price, hierarchy, images) alongside traditional traffic features. This multi-dimensional approach transforms the prediction model from relying solely on historical traffic data to integrating both structural and non-structural dimensions, thereby improving accuracy.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

2Reliability

If deep learning framework with multiple feature extractors is implemented, then comprehensive feature integration is achieved, but computing resources increase

Engineering Contradiction:

Improveprediction reliabilityVSAvoidcomputing resources

Core Design Contradiction:

ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent applies preliminary action by pre-processing and extracting meta-features (item-type, price, hierarchy, images) before the main prediction process. The multilayer perceptron integrates these pre-extracted features with traffic features, reducing the computational burden during real-time prediction and improving overall efficiency while maintaining reliability.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent uses copying by creating separate feature extractor components that independently process different data types (meta-features and traffic features). Each extractor produces a feature representation that is then combined, allowing parallel processing and efficient utilization of computing resources while achieving comprehensive feature integration.

Inventive Principle:

Principle #26Copying

3Adaptability or versatility

If non-structural information such as item-type, price, hierarchy, and image features is incorporated, then prediction coverage is enhanced, but data processing complexity increases

Engineering Contradiction:

Improveprediction coverageVSAvoiddata processing complexity

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements universality by designing a multilayer perceptron that universally processes both meta-feature representations and traffic feature representations. This single integration component handles multiple data types (item-type, price, hierarchy, images, and traffic features) through a unified architecture, enhancing prediction coverage while managing processing complexity through a standardized approach.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent introduces an intermediary multilayer perceptron that mediates between the meta-feature extractor and the prediction output. This intermediary component transforms diverse non-structural features into a unified representation that can be effectively combined with traffic features, simplifying the integration process and managing data processing complexity.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS12443973B2Deep learning-based revenue-per-click prediction model framework

Publication Date: 2025.10.14 WALMART APOLLO LLC

AI summary

A system including one or more processors and one or more non-transitory computer-readable media storing computing instructions that, when executed on the one or more processors, cause the one or more processors to perform operations including: reducing, using a multilayer perceptron, a dimension of feature embeddings for an item to generate a representation vector; extracting, using a recurrent neural network, sequential data from traffic features for the item; inputting the representation vector and the sequential data into a multilayer neural network, wherein the first scoring layer outputs one or more first performance metrics based on offline learning; and re-purposing the output of the second ReLU activation function of the multilayer neural network by feeding the output of the second ReLU activation function into a second scoring layer to output one or more second performance metrics based on online learning. Other embodiments are disclosed.