The application provides a kind of generation formula
shopping basket recommendation method based on multiple behavior sequence and special article embedding, belongs to the technical field of recommendation
system.The method of the application first extracts multiple interactive behavior records from user log, divides by time window and defines
behavior type as special article embedding
shopping basket sequence header, constructs heterogeneous
shopping basket sequence containing behavior Token;Then obtain article balanced semantic ID and complete multi-dimensional embedding superposition through residual quantization
autoencoder, generate input vector representation;Subsequently, extract fine-grained interaction features and shopping basket level long-term preference features through double-flow
encoder respectively, realize adaptive fusion through gate fusion unit;Then, predict the next dominant behavior based on the fused encoding features and use it as a prompt, combine with the location routing
hybrid expert network to generate article sequence in autoregressive manner;Finally, construct a multi-task joint
loss function to
train the model end-to-end, generate Top-K shopping basket recommendation
list and dynamically adjust it.