Research On Multi-Task Learning For Location Prediction From Trajectory Data

With the rapid development of mobile Internet and artificial intelligence,a large amount of mobile trajectory data is generated by various daily life applications.How to use these trajectory data to quickly and effectively predict the next location of users is extremely valuable,especially in location-based services(LBS)applications.In this thesis,the research goal is to improve the accuracy and efficiency of the location prediction using trajectory data,three methods,which use multi-tasking correlation,multi-model integration and model compression separately to predict location were studied.The next location of a moving object is often affected by other tasks.Existing location prediction methods typically use a single-task model without the shared features of multiple related tasks to improve each other's performance.Based on the mentioned facts,this thesis proposes a multi-task based location prediction model(MLoc),adopt a grid-based trajectory representation and then Embedding,and the obtained sequence vector is input into the long short-term memory network(LSTM),and then multiple tasks are simultaneously trained through the multi-task component,more meaningful feature representations are learned from the associated tasks,and the expressiveness of the model is enhanced.In this thesis,experiments are carried out on real datasets.The experimental results show that the accuracy of location prediction of MLoc model is superior to the ones of the current advanced models.In addition,the trajectory data has both a time dimension and a spatial dimension.Howerver,the existing model mainly considers the sequence feature of the trajectory,that is,the time feature.They often cannot accurately represent the spatial feature of the trajectory.That reduces the effect of the location prediction.This thesis proposes a multi-task and model-based location prediction model(MMLoc),it uses convolutional neural network(CNN)to extract spatial feature,which focuses on capturing the spatial relationship between the location and location of the moving object,and at the same time it uses LSTM to extract the location of the moving object sequence feature.Finally,the results of CNN and LSTM are integrated to form a multi-model component,and the features obtained by the multi-model component are input into the multi-task component to output the prediction result.The experimental results show that the Accuracy,Recall,Precision and F1-score indicators of the MMLoc model are superior to the ones of the current advanced models.Finally,considering the increasing application requirements of location prediction on the mobile devices,the location prediction model needs to work on resource-limited mobile devices.Therefore,how to compress the model to improve the efficiency of the model becomes an important research topic.In this thesis,the teacher model MMLoc is compressed into a simple student model by means of knowledge distillation.Specifically,by designing a multi-task distillation loss function,MMLoc is trained into a simple student model to achieve knowledge transfer.The experimental results show that the compression model can effectively reduce the size and improve the efficiency of the multi-task location prediction model.