Multi-behavior sequence recommendation method based on multi-granularity interest and application thereof

Abstract

The application relates to a multi-behavior sequence recommendation method based on multi-granularity interest and application thereof, collects multi-behavior historical interaction data of a user, forms a historical article interaction sequence, constructs a full connection graph and extracts a high-order multi-behavior dependent representation; a linear multi-head self-attention mechanism is used to extract a global behavior representation of the user and the multi-behavior dependent representation sequence is divided into sub-conversations, a multi-granularity multi-head self-attention mechanism is used to extract a multi-granularity user interest representation of all sub-conversations, the result is input into a nonlinear activation layer representation fusion, and a final embedding representation is obtained; a predicted score value of the user to an article at the position is obtained based on the final embedding representation and a candidate article embedding vector, and recommendation is carried out; and the application is applied to a user recommendation system. The application solves the data scarcity problem of a recommendation system, mines deep-level user interest expression and behavior patterns, solves the limitation that existing multi-behavior dependent representations are insufficient, reduces the calculation complexity, enriches user interest expression, and provides accurate personalized recommendation services.

Images