Research On Long And Short-term Sequential Recommendation Based On Deep Learning

In recent years,recommendation system has been a popular application in people's daily life.It can help people make fast and accurate choices in a short time.The traditional recommendation system did recommendations according to the long-term historical behavior and users' information.However,many users login anonymously in the actual process,which will not only lead to the cold start problem of the system,but also the actual prediction results not matching with the user's interest preferences.In this situation,the traditional recommendation system obviously can't keep pace with the needs of the social development.In order to deal with this problem,session-based sequential recommendation is put forward.In contrast to the traditional recommendation system,session-based sequential recommendation is based on the user's click sequences in the current session to predict,which can help capture the dynamic interests of user at any time.Although sequential recommendation performance has been improved greatly in contrast to traditional recommendation methods,there are still some problems: first,the time context features in the user session sequences and the higher-order features of the session sequences can't be effectively utilized.Secondly,if the long and short-term features are directly integrated to predict,the long and short-term features can't be effectively distinguished,and the time interval between session sequences,the categories of items in the session sequences and the position of items in the session sequences are not effectively utilized.In view of the above problems,our main work is as follows:(1)This thesis proposes a Time-aware Context and Feature Enhancement Sequential Recommendation(CAPNN)method.It can not only capture the high-order features of session sequences and time context features,but also can capture the long and short-term features of users.We first capture the short-term features from the user session sequences,and then capture the long-term features of the user.After that,a layer of attention mechanism is used to denoise the obtained long-term features,and the denoised feature vectors are fused with the point multiplication model to obtain the high-order features in the user session sequences.Finally,this thesis integrates the above features to make recommendations,the experimental results also verify the effectiveness of the model.(2)The above work does not distinguish the user's long and short-term interest features,which will cause the recommendation system to be unable to judge the influence of the user's long and short-term features on the final prediction result.Moreover,the time interval between the session sequences in the user behavior sequences,the items' categories and position features in the session sequences are not utilized.To this end,this thesis proposes a Multi-feature Aware Hierarchical Long and Short-term Sequential Recommendation(MFHLS)model.The model integrates the features of the time interval between the session sequences in the user behavior sequences,the category of the item in the session sequences,and the position where the item appears in the session sequences.And it distinguishes the long and short-term features of the user,using two layers of recurrent neural network to learn the user's long and short-term interest features separately,and finally combines the acquired long and short-term features to make recommendations,the experimental results also verify the effectiveness of the model.