Shared-Parameter Neural Networks for Unbiased Incremental Lift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional models for predicting incremental lift in online systems face challenges in processing large data volumes and produce biased results due to imbalanced treatment and control groups, making it difficult to accurately predict user actions on content items.
Innovation Solution
A machine-learned neural network architecture with shared and unshared parameters is used to generate control and treatment predictions, allowing for flexible modeling of user and item features, even with imbalanced data sets, by employing a neural network with a first portion for control predictions and a second portion for treatment predictions, sharing a subset of parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional models divide users into treatment and control groups for lift prediction, then lift evaluation can be performed, but the control group size is significantly different from the treatment group size leading to biased and inaccurate results
Solution Approach 1:
The neural network model is segmented into two distinct portions: a first portion that processes treatment group data and a second portion that processes control group data. This segmentation allows each portion to be optimized for its specific data type while maintaining balanced processing capabilities, thereby improving prediction accuracy and reducing bias from imbalanced group sizes.
Solution Approach 2:
Different portions of the neural network model are assigned different functions tailored to their specific data types. The first portion is optimized for treatment group characteristics while the second portion is optimized for control group characteristics, allowing each to process its respective data with appropriate local quality and reducing overall prediction bias.
2Quantity of substance
If conventional models process large amounts of user and item data, then comprehensive lift evaluation is possible, but the models have limited capacity to process large data volumes
Solution Approach 1:
The patent transforms the data processing approach by introducing a dual-portion neural network architecture that processes treatment and control group data in parallel through different pathways. This dimensional change in the model structure enables simultaneous processing of large volumes of both treatment and control data, significantly increasing overall data processing capacity while maintaining comprehensive evaluation capabilities.
3Device complexity
If a single neural network model processes both treatment and control predictions, then model complexity is reduced, but the model cannot differentiate between control and treatment predictions effectively
Solution Approach 1:
The neural network model is divided into two distinct portions with separate processing pathways for treatment and control predictions. This segmentation enables the model to maintain structural simplicity while effectively differentiating between prediction types through specialized processing in each portion, thereby preserving both simplicity and prediction accuracy.
Solution Approach 2:
Each portion of the neural network is assigned specific local quality characteristics optimized for its prediction type. The first portion is tuned for treatment prediction while the second portion is tuned for control prediction, allowing the model to differentiate effectively between the two while maintaining overall structural simplicity.
Data Source
AI summary
An online system trains a machine-learned lift prediction model configured as a neural network. The machine-learned lift prediction model can be used during the inference process to determine lift predictions for users and items associated with the online system. By configuring the lift prediction model as a neural network, the lift prediction model can capture and process information from users and items in various formats and more flexibly model users and items compared to existing methods. Moreover, the lift prediction model includes at least a first portion for generating control predictions and a second portion for generating treatment predictions, where the first portion and the second portion share a subset of parameters. The shared subset of parameters can capture information important for generating both control and treatment predictions even when the training data for a control group of users might be significantly smaller than that of the treatment group.