Research On Sequence Recommendation Algorithm Based On Deep Learnin

With the development of intelligent services on Internet platforms,it is crucial to create characteristic recommendation strategies for different users.Recommendation algorithms can reduce information overload and meet diverse business needs.However,existing recommendation algorithms ignore the dynamic changes of user preferences and lack research on sequential information when modeling user interest preferences.In addition,most current research focuses on improving the recommendation effectiveness of individual targets,ignoring the fact that user satisfaction and final decision are influenced by multiple targets.To address the above two problems,this thesis investigates the sequence recommendation algorithm based on deep learning,and the research work contains the sequence recommendation algorithm based on users’ long-and short-term interests and the sequence recommendation algorithm based on personalized multitasking.The details are as follows:(1)To address the problem that existing recommendation models only focus on the static correlation between users and items,ignoring the decay of users’ interests and the dynamic change of preferences,we design a Sequential Recommendation Model Based on User’s Long-term and Short-term Preferences(ULSPSR).First,we model users’ short-term interests and consider dynamic time intervals in the process of modeling short-term interest preferences,and we innovatively design a short-term attention network with time-awareness.Second,we model the long-term interests of users,and we innovatively design a multi-way null convolutional network in the long-term interest preference modeling process,and extend the network into a multi-pathway structure to mine the complex long-term behavioral features of users.Finally,considering that the importance of long-term and short-term interests changes dynamically,a long-and short-term interest-aware selector is proposed to adaptively match long-term and short-term interests.The ULSPSR model is experimented on Movie Lens-1M and Beauty datasets,and the experimental results show that the method can well reflect users’ preferences in different periods and improve the prediction accuracy.(2)To address the problem that existing recommendation models only model a single target task and ignore the fact that users’ interests are jointly influenced by multiple target tasks,we design a Sequential Recommendation Model Based on Personalized Multi-task(PMSR).On the one hand,in order to alleviate the negative migration phenomenon in the multi-task recommendation process,a personalized expert separation module is innovatively designed: first,the shared expert layer of the module uses the shared expert selection controller to mine the common association information among tasks;then,the exclusive expert layer of the module uses the exclusive expert selection controller to mine the unique implicit information within each task.On the other hand,in order to reduce the loss of high-level sequence association information,a personalized sequence knowledge transfer module is innovatively designed: first,the sequence interaction information within each task is mined at the task-specific tower network layer of this module;then,the influence of the previous task on the next task is considered at the sequence knowledge transfer network layer of this module to avoid the sparse data leading to poor recommendation results.Finally,the multi-task goals are combined for recommendation prediction.PMSR model is tested on two publicly available datasets,Ali-CPP and Tenrec,and the results show that the model outperforms other advanced benchmark models and achieves a more outstanding recommendation performance.