ZHAO Lulu , ZHAO Yuhong
(Digital and Intelligent Industry School , Inner Mongolia University of Science and Technology , Baotou 014010 , China)
Abstract : Dynamic time series is a key feature of many modern recommendation systems . Its primary aim is to capture the “context” of user activities based on their most recent actions . However , most LSTM-based sequence models only consider the user ’s short-term interests , neglecting their long-term interests . To enhance the performance of sequence recommendations , a Long Short-term Preference Recommendation Based on LSTM (LLSPRec) method is proposed here . This method models the user ’s time series with LSTM , aggre- gates relevant feature information between sequences to obtain the user ’s recent preferences , and models the distance between the user and candidate items using distance metric learning to capture the user ’s long-term preferences , dynamically integrating the user ’s long- term and short-term preferences according to their intentions , thereby accurately describing user interests and improving the diversity of recommendation results .
Key words : LSTM ; metric learning; dynamic time series; sequence recommendation; element correlation