Ranking Function for Ad Recommendation Using AUCL Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional binary classifier-based advertisement recommendation systems fail to accurately predict rare but important positive interactions, leading to a high accuracy in predicting negative interactions and neglecting potential user engagements due to the extreme class imbalance problem.
Innovation Solution
The use of ranking functions that learn ranking weight vectors by optimizing the area under the curve loss (AUCL) to compute scores for user interactions, prioritizing positive interactions and avoiding the class imbalance issue, with the scores being normalized using a threshold score to ensure accurate advertisement recommendations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If binary classifiers are trained to predict user interactions with advertisements, then the accuracy of predicting negative interactions is improved, but the ability to identify rare positive interactions deteriorates
Solution Approach 1:
The patent segments the prediction task into two separate models: a binary classifier for negative interactions and a ranking function for positive interactions. This segmentation allows each model to specialize in its respective task, with the ranking function specifically optimized to identify rare positive interactions by learning from both positive and negative examples and ranking them by likelihood.
Solution Approach 2:
The patent changes the optimization parameter from accuracy (which favors negative predictions) to area under the curve (AUC) for the ranking function. This parameter change enables the model to optimize for the ability to distinguish and rank positive interactions correctly, rather than simply maximizing overall prediction accuracy.
2Device complexity
If conventional binary classifiers are used to recommend advertisements, then the system complexity is reduced, but the user engagement rate deteriorates
Solution Approach 1:
The patent segments the recommendation system into a filtering stage (binary classifier) and a ranking stage (ranking function). This segmentation maintains relative system simplicity while significantly improving user engagement by ensuring that advertisements with rare positive interactions are properly identified and ranked.
Solution Approach 2:
The patent introduces a ranking function as an intermediary between the binary classifier and the final advertisement selection. This intermediary component processes the outputs of the binary classifier and re-ranks advertisements based on their likelihood of positive interaction, thereby improving engagement rates without substantially increasing overall system complexity.
3Reliability
If the ranking function optimizes for area under the curve loss, then the identification of positive interactions is improved, but the training complexity increases
Solution Approach 1:
The patent changes the training objective from accuracy optimization to area under the curve (AUC) optimization for the ranking function. While AUC optimization is more complex than accuracy optimization, it provides superior performance in identifying rare positive interactions. The increased training complexity is justified by the significant improvement in positive interaction prediction reliability.
Data Source
AI summary
A digital medium environment is described to recommend advertisements using ranking functions. A ranking function is configured to compute a score by applying a user context vector associated with a user to individual ranking weight vectors associated with advertisements, and provide the advertisement with the highest score to the user. In order to learn the ranking weight vectors for the ranking function, training data is obtained that includes user interactions with advertisements during previous sessions as well as user context vectors. The ranking weight vectors for the ranking function associated with each advertisement can then be learned by controlling the score generated by the ranking function to be higher for positive interactions than the negative interactions. To do so, the ranking weight vectors may be learned by optimizing an area under the curve ranking loss (AUCL) for the ranking function.