Design And Implementation Of User’s Point-of-Interest Prediction Algorithm

With the continuous development of information technology,an increasing number of location-based services have emerged on the Internet,such as intelligent transportation,logistics delivery,and unmanned driving.These services require predicting the future location of moving objects,such as vehicles or pedestrians,to optimize traffic,improve urban planning,and reduce accidents.As an important subtask in the location service,point-of-interest prediction(POI),by analyzing the user’s mobile mode,accurately predicting the user’s position at the next moment is of great significance for improving the location service and the user experience.However,existing methods for POI prediction suffer from certain limitations.For example,recurrent neural network-based prediction algorithms face the problems of "vanishing" or "exploding" gradients when dealing with long sequence data,which makes it difficult to capture longterm dependencies.Graph neural network-based methods are highly dependent on data quality.To address these limitations,we propose two models:the Long-and Short-term Preference Learning based on Prompt Embedding(LSPPE)and Multi-Task Learning for POI prediction based on Adversarial Enhancement(MTLAE).LSPPE uses attention mechanisms to learn a user’s long-and short-term preferences,allowing it to capture both local and global context information without sacrificing either.Additionally,the introduction of prompt embeddings and long-and shortterm fusion strategies further improves the model’s performance.MTLAE constructs POI features using graph representation learning,and employs three encoders to learn the user’s movement patterns for location,region,and category,respectively.The model then combines these features for POI prediction,and introduces adversarial samples to improve data diversity,mitigating the impact of noisy data.Furthermore,the introduction of Regularized Dropout for Neural Networks(R-drop)improves model training and inference consistency.Our experiments show that both proposed methods outperform several state-of-the-art algorithms.This thesis first introduces the current research status and related technologies for POI prediction at home and abroad,providing a theoretical foundation for readers.We then present the basic structure of the LSPPE model,and validate the model’s accuracy and the gains achieved by different improvements.Next,we introduce the basic structure of the MTLAE model,analyze its performance,and the effects of the proposed improvements.Finally,based on our proposed algorithms,we implement a user-oriented POI prediction prototype system.This system has complete trajectory processing,user information management,deep learning model inference,POI prediction,and data visualization frontend capabilities.