GENERAL NETWORK COMPRESSION FRAMEWORK AND COMPRESSION METHOD BASED ON SEQUENCE RECOMMENDATION SYSTEM

摘要

The present invention provides a general network compression framework and compression method based on a sequence recommendation system. The general network compression framework comprises: an input embedding layer based on block adaptive decomposition, used for dividing a set of recommended items into a plurality of clusters according to the frequency of the recommended items and dividing an input embedding matrix into a plurality of corresponding blocks, wherein different dimensions are assigned to the blocks of each cluster; an intermediate layer sharing layer parameters, the intermediate layer being connected to the input embedding layer, being formed by stacking a plurality of residual blocks, and sharing parameters by means of a layer parameter sharing mechanism; and an output layer based on block adaptive decomposition, the output layer using the same clustering configuration of block embedding as the input embedding layer, and representing the blocks of each cluster by using a tree structure, to obtain the probability distribution of an output sequence so as to predict an expected recommended item. The present invention effectively solves the problem of huge parameter quantity of a sequence recommendation model, improves the training and inference efficiency of the model, and alleviates the over-fitting phenomenon of the model.