To solve the problem of insufficient representation of session interest in bundle list recommendation and the problem that pre-built bundle cannot be personalized recommendation according to session dynamic interests,the global information of a session was extracted using a time-weight multi-head attention mechanism that incorporated relative positional encoding,the local information of the session was fused by the designed multi-granularity depth separable convolution network,and a diversified bundle list was generated through the autoregressive bundle list generation network.Extensive experiments were conducted on Amazon datasets and the size of generated bundle was analyzed.The model results are improved by an average of 15.19％com-pared to that of other optimal benchmark models.
维普
万方数据