Additive Decomposition Model for Long Attribution Window Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing solutions for content distribution in online systems are inadequate for predicting conversion events with a long attribution window, such as 14 days, which is crucial for content providers like those in the travel industry, as they often require longer time frames for user decisions, leading to inefficiencies in content delivery and resource allocation.
Innovation Solution
An online system employs an additive decomposition model with two trained models: one using fresh data for short-term conversion predictions within 1 day and another using delayed data for long-term predictions up to 14 days, ensuring accurate and up-to-date content selection based on user behavior.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If existing solutions use a 1-day attribution window for conversion prediction, then the prediction can be made quickly with available data, but the prediction accuracy deteriorates for long-term conversion events such as purchases that occur days or weeks after clicking
Solution Approach 1:
The patent segments the conversion prediction task into two separate models: a short-term conversion prediction model for immediate conversions (within 1 day) and a long-term conversion prediction model for delayed conversions (days to weeks). Each model is optimized for its specific time horizon, with the short-term model providing quick predictions and the long-term model capturing delayed conversion patterns. This segmentation resolves the contradiction by allowing speed-optimized short-term predictions while maintaining accuracy for long-term conversions through a dedicated model.
Solution Approach 2:
The patent implements a dynamic prediction system that adapts the attribution window based on the conversion event type and user behavior patterns. Rather than using a fixed 1-day window, the system dynamically adjusts the prediction horizon and selects appropriate models based on the specific conversion scenario. This dynamic approach allows the system to maintain both speed and accuracy by choosing the appropriate prediction strategy for each case.
2Measurement precision
If the online system waits to collect enough data before improving content delivery, then the conversion data becomes more accurate and complete, but the time delay increases, causing the system to miss optimal content delivery opportunities
Solution Approach 1:
The patent applies preliminary action by training and deploying the long-term conversion prediction model in advance, using historical conversion data that has already been collected. Instead of waiting for new data to accumulate before making predictions, the system proactively uses existing historical data to build predictive capabilities. This allows the system to immediately apply learned patterns to current content delivery decisions without waiting for additional data collection, thus resolving the contradiction between data accuracy and delivery timing.
Solution Approach 2:
The system implements a feedback mechanism where conversion outcomes (whether users eventually convert within the long-term window) are fed back into the model to continuously improve predictions. As more conversion data becomes available over time, the feedback loop refines the long-term prediction model, allowing the system to maintain accurate predictions without requiring extended waiting periods. The feedback enables iterative improvement while maintaining timely content delivery.
3Device complexity
If the online system uses a single conversion prediction model, then the system complexity is reduced, but the ability to accurately predict both short-term and long-term conversions deteriorates
Solution Approach 1:
The patent segments the conversion prediction functionality into two specialized models: a short-term conversion model and a long-term conversion model. Each model is designed and trained specifically for its time horizon, allowing both to achieve high accuracy without the compromises required by a single general-purpose model. The segmentation enables each model to focus on specific conversion patterns relevant to its time window, resolving the contradiction between simplicity and accuracy.
Solution Approach 2:
While using multiple models, the patent maintains a universal architecture where both short-term and long-term models share common feature extraction and prediction frameworks. This multi-functional design allows the system to handle different conversion time horizons through a unified approach, managing complexity through modular design while achieving comprehensive prediction accuracy across different time scales.
Data Source
AI summary
An online system optimizes for longer attribution window conversions with an additive decomposition model by predicting the probability that a predefined action happens given an impression/click. The online system receives a content item from a content provider for display to a target user, and predicts a probability that a target user will convert given an interaction with the content item by the target user. The online system computes, by a first trained model, a short-term conversion probability of a conversion event happening within a first conversion window after the interaction. The online system computes, by a second trained model, a long-term conversion probability of the a conversion event happening within a second conversion window after the interaction, the second conversion window being longer than the first conversion window. The online system computes the conversion probability given the interaction based on the short-term conversion probability and the long-term conversion probability.