Position Debiased Search Ranking via ML Noise Injection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional search systems suffer from position bias, where users tend to select top-ranked results regardless of their relevance, leading to inaccurate ranking of search results due to boosting low-positioned results, which may not be relevant to the user's query.
Innovation Solution
A position debiased search system is implemented using machine learning models, specifically deep neural networks, that train on position data with injected noise to rely on non-position features, and incorporates unseen results to avoid bias, ensuring that search results are ranked based on relevance rather than position.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If conventional search systems rank results by position, then users can quickly access top results, but position bias occurs where users select top results regardless of relevance
Solution Approach 1:
The system pre-trains a machine learning model using historical search data with position information, then applies this pre-trained model to new search queries. This preliminary training enables the system to automatically debias results based on learned patterns, resolving the contradiction between fast access and relevance accuracy without real-time computation overhead.
Solution Approach 2:
A machine learning model serves as an intermediary between the search query and result ranking. This model processes position information and other features to generate debiased rankings, acting as a mediator that transforms raw search data into relevance-based results while eliminating position bias automatically.
2Adaptability or versatility
If low-positioned results are boosted to reduce position bias, then more diverse results are shown, but irrelevant results may be over-ranked
Solution Approach 1:
The system changes the parameter used for ranking from position-based metrics to machine learning model predictions. By transforming the ranking parameter from simple position to a complex model output that considers multiple features, the system achieves both diversity and precision without manually boosting specific positions.
Solution Approach 2:
The patent replaces the mechanical boosting approach (manually adjusting ranks) with a machine learning-based system. This substitution allows the system to automatically learn which low-positioned results are truly relevant based on historical data, eliminating the need for arbitrary boosting while maintaining diversity and precision.
3Productivity
If machine learning models train on position data, then the system can learn from historical patterns, but the model may inherit position bias from training data
Solution Approach 1:
The system converts the harmful position bias present in training data into a beneficial feature. By explicitly including position information as an input feature to the model, it learns to recognize and compensate for position effects, ultimately producing debiased rankings that leverage historical patterns while eliminating bias.
Solution Approach 2:
Instead of trying to prevent position bias during training by excluding position data, the system inverts the approach by explicitly including position information and teaching the model to debias. This inversion allows the model to learn from position patterns while producing unbiased predictions, turning the potential harm into a learning opportunity.
Data Source
AI summary
A position debiased search system can avoid bias towards top-ranked search results using a position-trained machine-trained model. Past positions for listings can be input into the model with added noise and low-ranked results to train the model to generate rankings that do not exhibit position bias. A network site can implement the position debiased search system to generate network site results that can generate accurate user results in real time as users browse the network site.