Abstract: Traditional recommendation algorithms are mostly based on the static attributes of users without considering the drift of user’s interest. To this end, Top-N recommendation algorithm integrating user’s interest drift detection is proposed. The new algorithm makes full use of long short-term memory(LSTM) in processing time series data to represent short -term interest shift of users, compromises a fixed vector obtained by matrix factorization to represent long-term interest of users, and incorporates the attention mechanism into the representation of hidden state of the long short-term memory network to obtain the effect of the user’s long-term interest on the shortterm interest. Compared with current popular algorithms, performance of the proposed algorithm is superior in Top-N item recommendation.

Key words: recommendation algorithm, long short-term memory(LSTM), interest drift, matrix factorization, attention mech鄄 anism, time dynamics