Session-based recommendation method for enhancing multi-level interest perception through comparative learning

Abstract

The invention discloses a session-based recommendation method for enhancing multi-level interest perception through comparative learning, which comprises the following steps of: constructing a multi-level interest perception hypergraph channel, and learning different interest level item representations; constructing an intermediary interaction graph channel, and learning long-distance information between the items; splicing item feature information of the two channels through a fusion gating network, and learning session representations of the two channels by using a self-attention mechanism and a soft attention mechanism respectively; through comparative learning, mutual information between session embedding of the two channels is optimized; candidate item scores are calculated through repeated exploration standardization, a cross entropy loss function is used for learning, and comparison loss and cross entropy loss are unified into a learning target. The problems that an existing recommendation model based on sessions excessively pays attention to current interests of users represented by latest items in session sequences or interaction between single items, but cannot fully consider different levels of the interests of the users and sparsity of interaction sequences are solved.

Images