Brand Affinity Signal for Search Ranking Personalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional online shopping platforms fail to personalize search and browse rankings based on user brand affinity, leading to preferred brands not being prominently displayed, which can reduce user satisfaction and loyalty.
Innovation Solution
Implementing a system that calculates a brand affinity signal by combining product type prediction scores and user brand affinity scores, using an event-level learning-to-rank framework to rerank items on search and browse pages, and employing parallel processing to minimize latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional ranking methods are used based on relevance and sales performance, then items are ranked objectively according to general metrics, but user personalization and brand preference are not reflected
Solution Approach 1:
The patent segments the ranking system into multiple independent components: a base ranking model that handles objective metrics (relevance, sales performance) and a separate brand affinity model that handles user preferences. These segments are computed independently and then combined, allowing personalization without completely redesigning the ranking system.
Solution Approach 2:
The patent merges the base ranking score and brand affinity score through a linear combination formula: final_score = alpha * base_ranking_score + (1-alpha) * brand_affinity_score. This merging approach integrates personalization with objective metrics in a computationally efficient manner.
2Adaptability or versatility
If brand affinity signals are integrated into the ranking model, then user preference personalization is achieved, but computational complexity and processing time increase
Solution Approach 1:
The patent pre-computes brand affinity scores and stores them in a cache structure before they are needed for ranking. This preliminary action allows the ranking system to retrieve pre-computed affinity scores during query processing, avoiding the need to compute them in real-time and thus reducing latency.
Solution Approach 2:
The patent applies brand affinity signals selectively only to items that are relevant to the user's query and within the top-N results from the base ranking model. This local application focuses computational resources on the most impactful items rather than processing all items in the catalog.
3Ease of operation
If real-time brand affinity calculation is performed for each query, then personalized results are provided, but system latency increases
Solution Approach 1:
The system pre-calculates and caches brand affinity scores based on historical user behavior data before query time. During actual search queries, these pre-computed scores are retrieved from cache rather than recalculated, dramatically reducing the time required to provide personalized results.
Solution Approach 2:
The patent uses cached copies of brand affinity scores instead of recalculating them for each query. This copying approach allows the system to serve personalized rankings rapidly by reusing pre-computed affinity data that reflects user preferences over time.
Data Source
AI summary
A method including receiving a request from a user to view a page. The page is one of a search results page or a browse shelf page. The method also can include obtaining a respective brand affinity score for the user for each of one or more product types associated with the request. The method additionally can include generating a respective brand affinity signal for the user for each respective item in a baseline list of items to be displayed on the page, based on the request and the respective brand affinity score for the user for a product type of the one or more product types associated with the respective item. The method further can include generating a reranking of the items to be displayed on the page, based on a machine learning model and based on factors comprising the respective brand affinity signals for the user for the items and other rerank signals. The method additionally can include outputting the reranking of the items. Other embodiments are described.