Research On Destination Location Prediction Technology Based On BiLSTM

In recent years,with the extensive use of navigation and positioning technology,it has become possible to study the trend of target trajectory movement.Destination prediction based on the target historical trajectory plays a key role in the field of urban resources pre-allocation(taxi,shared bicycle,etc.)and accurate advertising recommendation.Driven by a large number of available trajectory data,destination prediction has always been a hot topic.Currently,the traditional methods such as Markov model and frequent pattern mining are mainly used for destination prediction.At the same time,with the rapid development of the machine learning,some algorithms based on Adaboost,MLP,LSTM have also been widely adopted and have achieved good research results.However,these methods are often sensitive to the sparse data and often could not effectively mine the deep temporal and spatial information behind the trajectories.It is still a big challenge to destination prediction based on the historical traj ector:y.Deep learning has been widely used in recent years,especially the Recurrent Neural Network(RNN)can model sequence data very well which the current output of a sequence is also related to the previous output.These advantages enable RNN to be applied naturally to the field of sequence information modeling and have achieved remarkable results.Based on the research status of destination prediction at home and abroad,we have proposed a destination prediction technology based on deep learning and embedding of spatiotemporal information.After the trajectory preprocessing,we firstly extracts the metadata features and the spatial-temporal factor matrix of the trajectory,and then predict the destination in an encoder and decoder manaer.The coding part uses a Bidirectional Long Short-Term Memory(BiLSTM)to learn the sequence features of the trajectory,and introduces the influence factors of spatial-temporal in the learning process.The decoding part adopts MLP which extracts the deep features of the trajectory sequence by taking the encoded partial output vector together with the attribute feature vector of the trajectory as input.At the same time,in order to improve the efficiency of model learning,the high-dimensional vector contained in the attribute features is reduced by the Word2vector algorithm.In this paper,the simulation test is carried out in the real taxi data set according to the above algorithm.The test results show that the target prediction error of the model reaches 2.53,and the prediction error under the dis@5 can reach 2.44,which outperforms the standard LSTM and BiLSTM models with more than 15%and 10%accuracies,respectively.The better prediction results also show that our algorithm proposed in this paper has higher feasibility and effectiveness.