Category Embedding Recommendations for Broader Item Discovery
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional recommendation systems are limited in scope, only providing related items based on user interactions and failing to suggest broader categories or families of items that the user may enjoy but cannot be determined from browsing history and feedback.
Innovation Solution
A category recommendation system that generates category embeddings from item embeddings and provides recommendations based on user and query embeddings, expanding suggestions to include whole families of items like American recipes or pub food recipes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional recommendation systems only provide related items based on user interactions, then the system complexity remains low and computation is simple, but the recommendation scope is limited and diversity is reduced
Solution Approach 1:
The patent segments the recommendation system into multiple independent components: item embedding generation module, category embedding generation module (that aggregates item embeddings), user embedding module, and query embedding module. This segmentation allows the system to handle complex tasks through modular operations, resolving the contradiction between expanded recommendation scope and system complexity by organizing complexity into manageable segments.
Solution Approach 2:
The patent introduces a new dimension (category level) above the traditional item level in the recommendation hierarchy. By generating category embeddings that aggregate multiple item embeddings, the system operates in an expanded dimensional space, enabling recommendations at both item and category levels simultaneously, thus increasing versatility without proportionally increasing complexity.
2Adaptability or versatility
If the system generates category embeddings by aggregating item embeddings for all candidate categories, then recommendation diversity improves, but computational time and processing resources increase
Solution Approach 1:
The patent implements preliminary action by pre-generating and storing category embeddings through aggregation of item embeddings during an offline phase. These pre-computed category embeddings are then reused during online serving, eliminating the need to re-aggregate embeddings for each recommendation query, thus reducing computational time while maintaining diversity.
Solution Approach 2:
The patent extracts only the essential category-level representations (category embeddings) from the full set of item embeddings, separating the computationally intensive aggregation process from the real-time recommendation process. This extraction allows the system to maintain diversity through category embeddings while reducing online computational burden.
3Adaptability or versatility
If the system uses both user embeddings and query embeddings for category selection, then recommendation personalization improves, but the complexity of embedding processing increases
Solution Approach 1:
The patent applies local quality by allowing different parts of the system to use different embedding combinations: user-specific recommendations use user embeddings, query-specific recommendations use query embeddings, and personalized query-based recommendations use both. This localized approach to embedding usage enables fine-grained personalization while managing complexity by applying the appropriate level of processing only where needed.
Data Source
AI summary
Techniques of providing category recommendations include a category recommendation system that provides users recommended categories based on implicit data (e.g., user-item interactions) and/or explicit data (e.g., queries, user information). The recommendations can be personalized or non-personalized (i.e., depending if user embeddings are used), queried or non-queried (i.e., depending on whether query embeddings are used), or personalized and queried (if both user and query embeddings are used). In any of these cases, there is an offline mode and a serving mode. In the offline mode, a category embedding is generated from an aggregation of item embeddings associated with a candidate category. In the serving mode, the candidate category is selected for display on a user device based on a similarity between the category embedding and either, or both, of the user embedding and the query embedding.