Deep Bayesian network modeling training method

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses a deep Bayesian network modeling training method, which comprises the following steps: preliminarily determining model nodes of a Bayesian network by adopting features of an analysis object, and determining the model nodes as deep nodes replaced by feature vectors obtained by a deep neural network; dividing the sample data into a training set and a test set to determine a deep neural network model structure and initialize network parameters, and inputting the original data into a deep neural network to obtain output features of the deep neural network, and obtaining a full sample training set after discrete quantization through the training set which is not discretized, obtaining the trained Bayesian network, obtaining the output characteristics of the Bayesian network, obtaining a full sample test set after discrete quantization, and achieving accuracy prediction of the output nodes of the deep neural network. According to the method, modeling reasoning can be carried out through the feature quantity obtained through feature engineering, and original data can be directly processed. The method is a Bayesian network model which has a complete mathematical theoretical basis, is good in interpretability and has an automatic feature extraction capability.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence, in particular to a deep Bayesian network modeling training method in the technical field of data mining. Background technique [0002] In recent years, the new generation of artificial intelligence technology represented by deep learning has continuously made breakthroughs in image recognition, speech recognition and other fields, which has greatly improved the performance of current machine learning algorithms. Data-driven deep learning methods can automatically extract features and abstract from data, and mine high-order correlations. However, the interpretability of such models is usually poor, and the ability to discriminate false information is not enough, and it is easy to commit critical issues. Wrong; at the same time, the deep learning model basically does not have real reasoning ability, nor can it combine the existing expert knowledge well. The dependence on a large amo...

AI Technical Summary

Problems solved by technology

However, Bayesian reasoning is generally based on feature quantities and does not have the ability to model and reason raw data. Its model structure, parameters and reasoning accuracy are severely restricted by feature selection.

[0004] Deep learning has strong automatic feature extraction capabilities. Bayesian networks have unique advantages in reasoning and introducing expert knowledge, but they do not have interpretability, poor robustness, are easy to be deceived, and rely on a large number of training samples; Bayesian networks The Si network can effectively combine expert knowledge and has a complete mathematical theoretical basis and good interpretability, but it can only model and reason the feature quantities obtained through feature engineering, and cannot directly process the original data.

Images