Heterogeneous Graph Neural Network for Search Personalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional retrieval systems face challenges in identifying relevant items quickly when users input only a few characters, leading to increased time and effort, and they often require lengthy queries due to inefficiencies in processing large numbers of candidate items, which consumes significant processor cycles and memory resources.
Innovation Solution
Training a heterogeneous graph neural network (GNN), such as GraphSAGE, to generate user and item embeddings that represent users and items, allowing for more accurate and efficient identification and ranking of relevant items based on user interactions and item metadata.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional retrieval systems process large numbers of candidate items, then comprehensive item coverage is achieved, but processing time and computational resources increase significantly
Solution Approach 1:
The system pre-generates user embeddings and item embeddings before actual retrieval operations. User embeddings capture user preferences from interaction history, while item embeddings represent item characteristics. During retrieval, these pre-computed embeddings enable rapid similarity computation without processing all candidate items from scratch, thus reducing processing time while maintaining comprehensive coverage.
Solution Approach 2:
Embeddings serve as an intermediary representation between raw user interactions and retrieval decisions. Instead of directly comparing user queries with all candidate items, the system maps both to embedding spaces where similarity can be efficiently computed. This intermediary layer enables scalable retrieval by transforming the problem into a geometric similarity search.
2Measurement precision
If conventional systems require lengthy queries, then retrieval accuracy improves, but user effort and time input increase
Solution Approach 1:
The system pre-computes user embeddings that encapsulate user preferences, browsing history, and interaction patterns. When a user submits even a partial query, the system combines this with the pre-computed user embedding to immediately generate accurate retrieval results. This eliminates the need for users to type lengthy queries, as the pre-processes have already prepared the necessary contextual understanding.
Solution Approach 2:
The system transforms the retrieval problem from text-matching based on query length to embedding-based similarity computation. By changing the parameter from query character count to embedding similarity score, the system achieves high retrieval accuracy regardless of query length, allowing users to input minimal characters while maintaining precision.
3Quantity of substance
If retrieval systems process all candidate items, then comprehensive results are provided, but computational resources and memory consumption increase
Solution Approach 1:
The system extracts only the essential features of users and items into compact embedding vectors. Instead of processing all raw interaction data and item metadata during each retrieval operation, the system pre-extracts key characteristics into low-dimensional embeddings. This extraction reduces the computational burden from processing full candidate item sets to computing similarities with pre-computed embeddings, significantly lowering processor cycles and memory usage.
Solution Approach 2:
The system creates embedding copies of user profiles and item characteristics that can be rapidly compared without accessing the original full datasets. These embedding copies serve as lightweight proxies that capture the essential retrieval information, enabling the system to evaluate many candidate items with minimal computational resources by working with compact vector representations rather than full item data.
Data Source
AI summary
Disclosed herein are system, apparatus, article of manufacture, method and/or computer program product embodiments, and/or combinations and sub-combinations thereof, for training a heterogenous graph neural network (GNN) to generate user embeddings corresponding to users and item embeddings corresponding to items. An example embodiment generates a first user interaction graph for a first time window and a second user interaction graph for a second time window, wherein each graph represents users and items as nodes and user-item interactions within the respective time window as edges, samples user-item node pairs from the second user interaction graph, and trains the heterogeneous GNN based on user-item node pairs from the first user interaction graph that correspond to the sampled user-item node pairs from the second user interaction graph. User and item embeddings generated by the trained GNN may be used to determine a relevancy of a given item with respect to a given user.