Autoencoder Auxiliary Features for Ad Prediction Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional methods for selecting winning ads in online advertising face challenges due to data sparsity, particularly because clicks are rare and conversions are even more so, leading to limited accuracy in performance prediction models.
Innovation Solution
The implementation of an enriched performance prediction model that utilizes auxiliary features generated by an autoencoder based on both click-attributed and non-click-attributed conversion data, which helps to address the data sparsity issue without impairing the original calibration of the model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional click-attributed conversion data is used for training performance prediction models, then model calibration is maintained, but prediction accuracy is limited due to data sparsity
Solution Approach 1:
The patent applies preliminary action by pre-training an autoencoder model on abundant non-click attributed conversion data before using it to generate auxiliary features for the main performance prediction model. This preliminary training phase allows the system to extract useful patterns from large amounts of unavailable data, which then serve as enhanced features to improve prediction accuracy in the main model without requiring the main model to be trained on sparse click-attributed data alone.
2Loss of information
If additional features are incorporated in learning the performance prediction model, then feature richness is improved, but data sparsity issue is not remedied and accuracy concern remains
Solution Approach 1:
The patent uses an autoencoder model as an intermediary to transform non-click attributed conversion data into auxiliary features. This intermediary component bridges the gap between abundant but insufficient data types and the performance prediction model that requires click-attributed data. The autoencoder extracts and compresses relevant information from non-click data, creating auxiliary features that enrich the input to the performance prediction model without directly adding more click-attributed conversion examples.
3Quantity of substance
If non-click attributed conversion data is leveraged for performance prediction, then data utilization is improved, but model calibration may be impaired
Solution Approach 1:
The patent segments the feature extraction and prediction tasks into two distinct components: an autoencoder model that processes non-click attributed conversion data to generate auxiliary features, and a separate performance prediction model that uses both original features and auxiliary features for final predictions. This segmentation allows each component to specialize - the autoencoder handles the abundant non-click data without needing to maintain calibration, while the performance prediction model maintains calibration by being trained on click-attributed data with enhanced feature input.
Data Source
AI summary
The present teaching relates to online advertising. Bids directed to a display ad opportunity are received, where the display ad opportunity involves a user and an associated context and each bid includes a candidate advertisement. Auxiliary features are obtained for each bid based on a code generated by an autoencoder based on the bid and a predicted performance metric is determined for the candidate advertisement associated with the bid based on the auxiliary features associated with the bid. A winning advertisement is selected from candidate advertisements of the bids according to a ranking determined based on the respective predicted performance metrics of the candidate advertisements.