Research On Taxi Travel Destination Prediction Technology Based On Multi-scale Convolution Neural Network

In recent years,the development of mobile communication and navigation and positioning technology has made location-based service(Location Based Service,LBS)became an important part of people's daily life.With the rise of Location information service(Internet plus location information),it is more valuable to study the trend of target trajectory movement for destination prediction.For example,predicting a user's travel destination can optimize resource scheduling,Discover hidden user behavior preferences in the location,and achieve targeted ad push,popular location recommendations.At present the destination prediction algorithm commonly based on historical travel trajectory,such as frequent pattern Mining and Markov.Thanks to the development of parallel computing and the rapid rise of deep learning,a new breakthrough has been made in the destination prediction in response to a large number of complex and diverse trajectory data.Many machine learning models,such as Multi-Layer Perceptron,Recurrent Neural Network,have been widely used in the field of position prediction.However,due to the characteristics that of traj ectory data uneven distribution,long sequences,uneven density,sparse and semantic loss.The current research results still have the problem of insufficient utilization,lack of association between different moving objects of data features.Recent years,the emergence of Convolutional Neural Networks(CNN)has pushed deep learning to the forefront of most machine learning tasks.Because the classical CNN models have lower adaptability of scale,many researchers have designed suitable convolutional neural networks for trajectory Mining,the designed models mainly focuse on improving the ability of feature learning.It makes CNN widely used in destination prediction technology and achieving great results.Based on the existing researches of location prediction,this paper proposes a destination prediction method based on deep learning and attention mechanism,and constructs a new destination prediction model:MSCNN.After the trajectory preprocessing,the ID and user information metadata are embedded and then fed to FM,the FM can automatically extracts the second-order combination features from input and then the outputs are fed to the multi-layer perceptron network.The model can fully learn the information outside the GPS data;As for the CNN part,the trajectory data is randomly truncated,and then the trajectories are transformed to grids and pixels.Aiming at the multi-scale characteristics and density difference of trajectory data,the convolution layer composed with multi-scale convolution kernels,and the visual attention mechanism is introduced,the self-learning weights are obtained to re-scale the feature map.The attention mechanism can enhance the ability of spatial attention,and Mine the deep features of the trajectory matrix;the multi-layer perceptron and convolutional neural network are joint trained,the travel destination is generated by weighting the Mean-Shift hotspot.This paper uses the real taxi trajectory data of Porto,Portugal to conduct experiments.After the random truncation and cluster analysis,the destination of taxi in the urban area is predicted according to trajectory prefix,and the prediction result is calculated by the Haversine distance function.The distance loss of proposed model is about 2.5 kilometers,it is significantly better than the 2.75 km of the traditional neural network and 2.55 km of the basic convolutional neural network.The result proves the validity and reliability of MSCNN.The multi-scale convolutional neural network proposed in this paper provides a new method for destination prediction research,which has good application value and development prospects in practical scenarios.