Fusion Ordering Model for Search Result Ranking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current search engines face challenges in balancing multiple user requirements and optimizing search ordering results to meet traffic and income goals, leading to suboptimal user experience and monetization efficiency.
Innovation Solution
A fusion ordering model is trained using prediction scores from candidate search results, with feedback information collected and updated through a combined function, to adaptively optimize search result ordering, balancing traffic and income targets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a search engine uses traditional ordering methods to balance user requirements and monetization goals, then the system structure remains simple, but the ordering results become suboptimal and fail to meet traffic and income goals
Solution Approach 1:
The patent segments the ordering problem into multiple independent target models (e.g., traffic model, income model, user experience model), each optimizing for a specific goal. These segmented models then feed into a fusion ordering model that combines their outputs, allowing the system to achieve complex multi-objective optimization without creating an monolithic complex model structure.
Solution Approach 2:
The patent merges multiple prediction scores from different target models into a unified fusion ordering result. By combining the outputs of specialized models (traffic, income, user experience) into a single fused ordering, the system achieves comprehensive optimization across multiple goals while maintaining the simplicity of individual component models.
2Adaptability or versatility
If the search engine uses static ordering models, then the model structure remains simple, but the system cannot adapt to changing user interactions and preferences
Solution Approach 1:
The patent implements a feedback mechanism where actual user interaction data (clicks, dwell time, conversions) is collected and used to retrain and update the target models and fusion ordering model. This closed-loop feedback system enables the models to continuously adapt to changing user preferences and behaviors, transforming a static system into a dynamic adaptive one.
Solution Approach 2:
The patent transitions from static ordering models to dynamic models that can adapt their parameters and structures based on real-world performance. Through iterative training with feedback data, the models evolve to reflect current user behaviors and preferences, enabling the system to dynamically adjust to changing conditions.
3Productivity
If the search engine optimizes for multiple targets simultaneously using traditional methods, then the system attempts to meet multiple goals, but the ordering results become suboptimal and fail to balance traffic and income effectively
Solution Approach 1:
The patent divides the multi-target optimization problem into separate single-target optimization subproblems. Each target model (traffic, income, user experience) is trained to optimize its specific goal independently, achieving high precision for each individual target. The fusion ordering model then combines these specialized optimizations to achieve overall multi-target optimization with maintained precision.
Solution Approach 2:
The patent changes the optimization parameters by training separate models for different targets rather than attempting to optimize all targets simultaneously with a single model. This parameter separation allows each model to focus on its specific optimization goal, improving the precision of ordering results for each target while achieving comprehensive multi-target optimization through fusion.
Data Source
AI summary
A method for training a fusion ordering model, a search ordering method, an electronic device and a storage medium, and related to the technical field of artificial intelligence such as deep learning and the like are provided. The method includes: inputting prediction scores of a plurality of targets included in a candidate search result into a fusion ordering model, to obtain a fusion ordering result; collecting feedback information of the plurality of targets included in the fusion ordering result; and updating the fusion ordering model by utilizing the feedback information and a combined function, wherein the combined function is a function constructed by utilizing the plurality of targets. The solution of the embodiment is favorable for balancing the requirements of the plurality of targets, and the obtained updated fusion ordering model can provide a better search ordering result.