Development Of A Model For Location Prediction Based On Trajectory Data

With the rapid development of mobile Internet,positioning technology and the popularity of smart portable devices,artificial intelligence-led technologies and applications are changing people's lives.Among them,location-based services are one of the important technical supports.Location prediction is an important part of location-based services and plays an important role in areas such as recommendation systems and urban resource planning.Currently,Global Positioning System(GPS)based trajectory data has received wide attention in location prediction tasks.GPS trajectory data is spatio-temporal sequence data,which not only contains time and location information,but also contains rich contextual information in the trajectory sequence.Most of the traditional location prediction methods focus on the location sequence in GPS trajectories,without fully exploiting the contextual information in the trajectories,resulting in poor prediction results.One of them is that travel mode is one of the important contextual information that affects the location prediction results.The existing studies do not pay sufficient attention to the impact of the user's travel mode in the historical trajectory on the location prediction results,which leads to the modeling and prediction of mixed travel mode trajectories,and the generalization ability of the model is insufficient without the support of enough training data.Secondly,different locations in the historical trajectory are not enough to predict the location.Second,the weights of different locations in the historical trajectories are different for the prediction results,and the existing studies do not focus on the locations in the historical trajectories that have a greater impact on the location prediction results,which reduces the model prediction accuracy and training efficiency.Thirdly,most of the existing studies extracted the stay locations or interest points from the historical trajectories,thus transforming the location prediction problem into a classification problem,which is a disguised simplification of the location prediction problem,resulting in the prediction results being limited by the locations reached in history.To address the above problems,this thesis proposes a location prediction model combining travel modes based on real GPS trajectory data.The model firstly segments the trajectories according to the time threshold and carries out pre-processing and feature extraction;then identifies the travel modes in the trajectories by the improved deep forest algorithm,classifies the trajectories by different travel modes according to the travel mode identification results,and builds the corresponding location prediction model for different travel modes;finally,it adopts a Bi LSTM(Bi-directional Long Short-Term Memory)model with the attention mechanism(Attention)for location prediction.The attention mechanism in the model and the bi-directional propagation mechanism in the LSTM are used to explore the influence of different locations in the historical trajectory on the prediction results and improve the prediction accuracy of the model.The final prediction result of the model is the actual location that the user will reach,so that the prediction result is not limited to the location that the user has reached in the historical trajectory.Comparative experiments on real GPS trajectory datasets show that the improved deep forest-based travel mode identification model proposed in this thesis achieves better results in the travel mode identification task with an accuracy of 81.56%,which is better than the random forest models,and can be used as the basis for the location prediction model in this thesis to provide trajectory classification.The location prediction model proposed in this thesis classifies the original trajectory segments by different travel modes with the assistance of travel mode identification task,which improves the overall generalization ability of the location prediction model,and its accuracy improves 12.8% compared with the model without combining travel modes,which indicates that the model and algorithm proposed in this thesis have high feasibility and effectiveness.