Research On Algorithms For Multi-tasking Problems In Personalized Recommendation Scenarios

Due to the diversity of user behaviors,it is not comprehensive to portray user behaviors with only a single task,and traditional methods to solve multi-task problems are more likely to build multiple single-task models to predict multiple tasks,which requires more computational resources and may lead to suboptimal results due to the lack of information about the correlation between tasks.Therefore,in this thesis,we combine real anime recommendation data with the problem of inconsistent evaluation criteria and ambiguity in evaluation,and use the similarity of different subjects on different tasks and the correlation between different tasks to study the feature construction based on graph representation learning and the multi-task learning algorithm based on HMM in the anime recommendation scenario.The contributions of this thesis are summarized as follows:(1)Based on the similarity of different subjects on different tasks in the animation recommendation scenario,this thesis constructs a heterogeneous graph with weights between subjects and tasks,and extracts information from the subject nodes within the heterogeneous graph by using graph representation learning algorithms,and the experimental analysis shows that this information can effectively improve the performance of the model.(2)Based on the existence of correlation between rating task and status task,a multi-task learning model of HMM-PLE is proposed,which explicitly constructs the initial status probability vector,the status transfer probability matrix and the status-to-rating observation probability matrix on the basis of the PLE model,and learns them as subtasks.Experimental analysis compares the traditional multi-model fusion model and other multi-task learning models,and the results show that the HMM-PLE model outperforms the other models.