Deep-Embedding Architecture for Personalized Item Recommendations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems for providing personalized item recommendations in online grocery shopping struggle to effectively capture user preferences and item complementarity, leading to suboptimal recommendations due to limitations in handling large baskets and high latency in inference processes.
Innovation Solution
A large-scale deep-embedding architecture using a triple embeddings model, which trains on user, item, and basket interactions to generate personalized recommendations by leveraging a skip-gram framework and approximate nearest neighbor indexing for efficient inference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional recommendation systems are used to capture user preferences and item complementarity, then personalization can be achieved, but the system suffers from high latency in inference processes and suboptimal recommendations when handling large baskets
Solution Approach 1:
The system performs preliminary actions by pre-computing dense embedding vectors for all items in the catalog during an offline training phase using the triple embeddings model. These embeddings capture user preferences and item complementarity in advance, transforming the online inference problem into a simple similarity search that can be executed with minimal latency
Solution Approach 2:
The patent replaces the traditional mechanical recommendation system that processes user baskets and computes recommendations in real-time with a data-driven embedding-based system. By substituting the complex inference mechanism with pre-computed embeddings and similarity search, the system achieves both accuracy and speed
2Adaptability or versatility
If traditional recommendation systems process large baskets, then comprehensive recommendations can be provided, but the system performance degrades due to computational complexity
Solution Approach 1:
The system pre-computes dense embedding vectors for all items in the catalog during an offline training phase using the triple embeddings model. These embeddings capture user preferences and item complementarity in advance, transforming the online inference problem into a simple similarity search that can be executed with minimal latency
Solution Approach 2:
The patent changes the parameter representation from discrete item categories and rules to continuous dense embedding vectors. This transformation allows the system to handle arbitrary basket sizes efficiently by computing similarity in the embedding space, where computational complexity becomes independent of basket size
3Adaptability or versatility
If personalized recommendations are generated for millions of users, then user-specific preferences can be captured, but the system complexity increases significantly
Solution Approach 1:
The patent creates a universal embedding space where all users, items, and baskets are represented in a common vector space. The triple embeddings model serves multiple functions simultaneously: it captures user preferences, item characteristics, and basket-level complementarity relationships, eliminating the need for separate processing systems for each function
Solution Approach 2:
The system transforms complex user preference data into compact dense embedding vectors that capture essential characteristics in a low-dimensional space. This parameter transformation simplifies the system architecture by replacing complex rule-based personalization logic with straightforward vector similarity computations
Data Source
AI summary
A method including receiving a basket including basket items selected by a user from an item catalog. The method also can include grouping the basket items of the basket into categories based on a respective item category of each of the basket items. The method additionally can include randomly sampling a respective anchor item from each of the categories. The method further can include generating a respective list of complementary items for the respective anchor item for the each of the categories based on a respective score for each of the complementary items generated using two sets of trained item embeddings for items in the item catalog and using trained user embeddings for the user. The two sets of trained item embeddings and the trained user embeddings can be trained using a triple embeddings model with triplets. The triplets each can include a respective first user of users, a respective first item from the item catalog, and a respective second item from the item catalog, in which the respective first user selected the respective first item and the respective second item in a respective same basket. The method additionally can include building a list of personalized recommended items for the user based on the respective lists of the complementary items for the categories. The method further can include sending instructions to display, to the user on a user interface of a user device, at least a portion of the list of personalized recommended items. Other embodiments are disclosed.