Research And Implementation Of Automobile Travel Destination Prediction Based On XGBoost

In recent years,with the continuous development of new business models such as shared bicycles and online car-hailing,research on user travel behavior based on user travel data can help provide decision support in different applications.For applications like Didi Chuxing,if the user can accurately predict the intended destination when the user opens the APP,the user experience can be greatly enhanced.At the same time,it can also help solve the problem of order scheduling and demand forecasting,improve the rate of online carhailing and reduce the waiting time for users.Most of the existing purpose prediction studies are based on historical travel trajectory data.The methods used for destination prediction are mainly divided into two categories: statistical methods and neural networks.Statistical methods mainly use Markov models.Such methods often use external information to improve the accuracy of predictions.The neural network method has the following advantages: it converts trajectory data into trajectory images to ensure the spatiotemporal relationship between trajectory data;and it can automatically extract advanced features from trajectory images through deep learning.However,due to the sparseness of the trajectory data,the transformed trajectory image is inaccurate,and there is a lot of noise in the trajectory image.This article is based on real car trip data.Given historical travel records,the destination prediction method is used to convert the destination prediction to the probability that the user will go to each candidate destination in the candidate destination set.It does not rely on the user’s initial trajectory data,and performs destination prediction based on departure time,departure longitude,and latitude.Destination prediction can be performed when the user is preparing to travel,and the validity and feasibility of the method are verified by experiments.The main work of this article is as follows:(1)After coding latitude and longitude by geohash,there are many destination areas that need to be predicted.Direct prediction will bring huge overhead,and candidate destination screening is required.Since it is to screen out possible destinations,no precise answer is needed,and simple and reasonable rules can be used to screen out suitable candidate destinations.First,select some reasonable candidate destinations by simple and reasonable rules,so as to reduce the training volume of subsequent models and improve the data utilization rate.Then,for these candidate endpoints,statistical features,temporal features,neighbor features,and graph theory features are extracted from three perspectives: user,time,and location.These features are then input into the XGBoost model for training.The trained model is used to predict the probability of each candidate destination in the candidate destination set.After the prediction probability is sorted from large to small,the candidate destination with the highest probability is selected as the prediction result.(2)Based on the destination prediction model trained in the experiment,this paper designs and implements a destination prediction system with a simple and easy-to-use interface,easy to expand and maintain.When the user enters the system,according to the current time and the latitude and longitude of the user’s starting position,the model of the experiment is used to predict the travel destination,and the most likely first three destinations are recommended to the user as the prediction result display.