Multi-sequence Attention Decoupling Fusion Method for Sequential Recommendation
Sequential recommendation methods based on the attention mechanism in the recom-mendation domain have achieved good performance,but there still exists the phenomenon of premature embedding of auxiliary information,which leads to the following problems:1)the embedding matrix is too large,which increases the complexity of the attention computation;2)the heterogeneous information from different sources is fused into the same embedding matrix for the attention computation,which makes the model unable to distinguish the heterogeneous infor-mation from different sources;3)the embedding matrix's gradient is relatively fixed and cannot be flexibly transformed according to the application scenario.To this end,a multiple-sequence attention decoupling fusion method for sequence recommendation(MADF-SR)is proposed.First,to reduce the computational complexity,MADF-SR performs the attention computation of item sequences and auxiliary information sequences(e.g.,time-context sequences and category-context sequences)separately.Also the heterogeneous information from different sources is subjected to attention computation independently,which also enables the model to better distinguish hetero-geneous information.Then,MADF-SR selectively fuses the attention matrix obtained from the attention computa-tion with the item sequences according to the needs of the application scenarios,so that the embedding matrix can adaptively change the gradient according to the changes in the number of selected auxiliary information sequences.Experimental results on four real datasets show that MADF-SR performs significantly better than the baseline model on the evaluation metric NDCG@10.
NETL
NSTL
万方数据
维普
