Traffic accident rate predicting system based on online variational Bayesian support vector regression

Abstract

The invention relates to a traffic accident rate prediction system based on on-line variational Bayesian support vector regression, which comprises a data preprocessing module, an online variational Bayesian support vector regression model building module, an online variational bayesian support vector regression model training module, and an online variational bayesian support vector regression model prediction module. This method effectively solves the problem that the traditional support vector regression model predicts the speed of traffic accident rate is slow, the prediction result is inaccurate, and it is difficult to solve the problem on line and show its practical value.

Description

technical field [0001] The invention relates to a high-efficiency traffic accident rate prediction system based on online variational Bayesian support vector regression, which belongs to the application field of machine learning in traffic. Background technique [0002] The traffic accident rate prediction problem is essentially a regression problem. A classic way to solve this problem is to use the traditional support vector regression model, which has better generalization ability by compromising experience loss and model complexity. Although widely used, its main variants are based on the MAP criterion, which estimates a value for each parameter of the model, so it is more susceptible to interference from noise and outliers. To make matters worse, this model requires the user to specify the regularization parameters in advance, but in the application of many machine learning problems, it is difficult to determine the optimal regularization parameters in advance, and unre...

AI Technical Summary

Problems solved by technology

However, both works lack smoothing for ∈-insensitive losses, and it is difficult to give precise solutions to the parameters of the model due to the direct use of non-differentiable loss functions, thus adopting a compromise approach that limits the prediction of the model ability

In particular, due to the relatively few features of the traffic accident rate, the predictive ability of the model is more limited

Since then, Ning et al. proposed a robust Bayesian support vector regression model based on minimizing the square loss, and transformed the inequality constraints into equality constraints to deal with outliers, but this loss function is differentiable, not ∈ - insensitive loss, thus slightly insufficient in guaranteeing the sparsity of the model

This will cause the model to be more memorized than learned, and it is difficult to give reasonable predictions when the combination of features related to the accident rate does not appear in the training data

[0003] On the other hand, the traffic accident rate prediction problem is a streaming data problem

If you do not consider the actual environment, only learn a specific model based on historical data, but cannot integrate new data into this model in real time, then such a model is meaningless and cannot be put into practical application

However, many existing online support vector regression models are based on point estimation without considering the uncertainty of the model

This leads to its predictive performance being easily affected by unreasonable regularization parameter estimates, noise and outliers

Ma et al. proposed an accurate online support vector regression model, however, this method requires an uncertain number of operations when updating the model, which is very time-consuming, so it is not suitable for real-time environments

Then, Kivinen et al. proposed a method based on stochastic gradient descent to give an approximate solution in a limited time, but the predictive performance of the model is related to the initialization and learning rate of stochastic gradient descent, so the predictive ability of the model cannot be guaranteed

Brugger learns the model by weighing the accuracy and efficiency of the original optimization problem, but usually the problem needs to ensure the prediction accuracy and consume as little time as possible, and Brugger's analytical solution method is difficult to achieve these two goals at the same time

Images