Recommendation Reranking Model for Object Sequence Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for ranking target objects to maximize user feedback, such as click-through rate and conversion rate, fail to consider the influence of multiple objects shown together, leading to suboptimal user engagement and platform revenue in recommendation systems.
Innovation Solution
A method involving an initial ranking of target objects based on correlation degrees with the user, followed by a reranking process using a pre-trained model that combines attribute features, correlation degrees, and user features, employing encoders like transformer networks or RNNs to predict preference degrees and reorder objects for improved user interaction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If target objects are ranked in descending order of correlation degrees with the target user, then the ranking process is simple and fast, but user feedback such as click-through rate and conversion rate cannot be maximized
Solution Approach 1:
The ranking process is divided into two independent stages: initial ranking based on correlation degrees and reranking based on mutual influence. The initial ranking stage quickly sorts objects by correlation degree, while the reranking stage adjusts positions considering mutual influence among objects. This segmentation allows each stage to focus on specific objectives without compromising overall performance.
Solution Approach 2:
The initial ranking based on correlation degrees is performed as a preliminary action before the reranking process. This preliminary ranking provides a baseline ordering that is then refined by considering mutual influence among objects. The preliminary action ensures that the most relevant objects are identified first, and subsequent adjustments optimize user feedback.
2Device complexity
If only correlation degrees are used for ranking, then the ranking process is simple, but the mutual influence among multiple objects shown together is not considered
Solution Approach 1:
The ranking process merges two types of information: correlation degrees between objects and user, and mutual influence among objects. The initial ranking provides ordering based on correlation degrees, and the reranking process integrates mutual influence information to adjust positions. This merging ensures comprehensive consideration of both individual relevance and contextual relationships.
Solution Approach 2:
The reranking process acts as an intermediary that bridges the initial correlation-based ranking and the final presentation to the user. It takes the initial ranking as input and transforms it by incorporating mutual influence information, producing an optimized sequence that maximizes user feedback while maintaining the foundational relevance established in the initial ranking.
3Reliability
If a reranking process considering mutual influence among objects is implemented, then user feedback is maximized, but the system complexity increases
Solution Approach 1:
The system is segmented into two distinct modules: an initial ranking module that handles correlation degree calculations and a reranking module that processes mutual influence. This segmentation allows each module to be optimized independently, with the initial ranking providing efficient baseline ordering and the reranking module focusing specifically on adjusting positions based on mutual influence without duplicating computational efforts.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on computer storage media, are provided for displaying a target object sequence to a target user. One of the methods includes determining an initial ranking of a plurality of target objects in a target object sequence. A plurality of object features corresponding to the plurality of target objects are obtained. The plurality of object features are input, based on the initial ranking, into a reranking model. The plurality of object features are encoded to obtain a context representation vector corresponding to the target object sequence. A reranking of the plurality of target objects is obtained based on the context representation vector. The target object sequence is displayed to the target user based on the reranking of the plurality of target objects.