Retrieval Relation Graph for Search Intent Inference

The patent transitions from traditional keyword-based recommendation to a graph neural network approach that models user search behavior as a retrieval relation graph. This dimensional shift from flat keyword matching to multi-dimensional relational modeling captures the semantic relationships and user intent more accurately, reducing recommendation errors by considering the structure and context of search queries rather than just keyword frequency and correlation.

The patent introduces a graph neural network as an intermediary layer between keyword extraction and recommendation generation. This intermediary processes the raw search queries through learned semantic representations, transforming inconsistent keywords into accurate intent representations. The GNN model acts as a mediator that reconciles the efficiency of keyword processing with the accuracy of intention capture by learning from historical search behavior patterns.

The patent replaces the mechanical weighted summation system with a neural network-based learning system. Instead of manually defining weights for search frequency and correlation, the graph neural network automatically learns the appropriate weighting and relationships from historical data. This substitution of mechanical calculation with intelligent learning enables the system to handle complex user intentions while maintaining reasonable computational complexity through distributed representation and parallel processing.