Abstract: Aiming at the problem that the existing sequence recommendation algorithms did not make full use of time information, we proposed  a multi-time embedding mode,  which used  dynamic fusion strategy to alleviate the problem of insufficient long-term preference modeling in existing sequence recommendation algorithms. The multi-time embedding mode could simultaneously model the absolute time information and relative time information of user-item interaction, and fully capture various rules of user-item interaction about time. The dynamic fusion network dynamically integrated the users’ long-term preference and recent preference according to the users’ intentions,  accurately depicted the users’ interest, and improved the diversity of recommendation results.  The proposed time embedding sequence recommendation algorithm based on  long-term memory enhancement was compared with the existing  algorithms on the public datasets MovieLens-1M and Amazon-Beauty. The results show that the proposed algorithm is better than the comparison method in the evaluation indicators HR@N and NDCG@N. Experimental results show that the accuracy of the time embedding sequence recommendation algorithm based on  long-term memory enhancement is higher than that of other  comparison sequence recommendation algorithms.

Key words: deep learning, recommendation algorithm, attention mechanism, time embedding, long-termand short-term interest