Research On Recommendation Algorithm Based On Sequence Feature Extraction

In recent years,sequential recommendation systems have been widely studied and applied in industry and academia due to their ability of capturing users' long-term or short-term interest preferences by combining temporal information,surpassing the traditional collaborative filtering-based recommendation methods.In general,according to whether the sequential recommendation system interacts with users or not,it can be divided into static and dynamic:1)static sequential recommendation systems focus on mining users'historical interaction information to construct users' interests,and are oriented to short-term recommendations;2)dynamic sequential recommendation systems not only exploit historical interaction information,but also model users' feedback information in the rec-ommendation process,which is used to track users' interests in real time,and are more oriented to long-term recommendations.With the development of deep learning,both academia and industry are committed to applying deep learning to more complex sequen-tial recommendation tasks to improve their prediction accuracy.However,existing deep sequential recommendation models have some shortcomings:lack of careful design of user presentation,neglect of filtering invalid or noisy information and inaccurate model-ing of users' dynamic interests.Therefore,the research in this thesis is devoted to con-structing more effective static/dynamic sequential recommendation models,and the main work includes:(1)Introducing information entropy,this thesis proposes a self-attentive static se-quential recommendation model based on entropy regularization.It has obvious differen-tiation for different items,better capture the items that characterize the user's interest,and reduce the adverse effect of noisy items on constructing the user's interest.(2)A new residual connection is applied to construct a residual self-attentive static sequential recommendation model.It compensates for the disadvantages of unstable train-ing,sensitivity to parameters,and the necessity of preheating learning rate strategy in the traditional static sequential recommendation model based on the self-attention mecha-nism.(3)A single-queue user modeling approach based on positive and negative feedback is constructed to improve the traditional dual-queue modeling method.It can more clearly represent the user's preference,more accurately dynamic model the user's interest state and reduce the number of parameters required by the model.(4)Introducing self-attention mechanism,this thesis proposes a self-attentive dy-namic deep reinforcement learning recommendation model.It can better obtain the the user's interest representation features,dynamically capture changes in user interests,and better solve the user cold-start problem in traditional recommendations.(5)The effectiveness of the proposed different approaches is explored and validated through extensive experiments in different offline data sets under the static and dynamic sequential recommendation scenarios.Different perspectives on the factors affecting user interest modeling in sequential recommendation systems are provided for discussion and validation.