Neural network algorithm for processing structured data

Abstract

The invention discloses a neural network algorithm for processing structured data. The neural network algorithm comprises the following steps: creating a Transform structure neural network with multiple activation function selections; adding a Rezero mechanism on the basis of jump connection of a Transform structure neural network to obtain a Rezero mechanism Transform structure neural network; adding a FeedForward structure to a part output by the Rezero mechanism Transform structure neural network, and obtaining a structured data neural network model; and adopting the structured data neural network model to carry out cycle training on the data, and completing the cycle when a training effect is achieved.

Description

Technical field [0001] The present invention is in computer science depth study, based on a model of the structured data field, particularly to a neural network algorithm for processing structured data. Background technique [0002] In the era of data-driven, deep learning algorithm has received extensive attention, but at the same time, machine learning is also booming, with the depth of deep learning neural network model represented, and machine learning model tree as the representative; despite the depth of the neural network made in computer vision and natural language processing was a great success, but in many traditional areas, such as wind control model, based on the task of fraud and other forms of data, the performance is not good; recently, Google made in the form of data for the neural network model TabNet, by test illustrates popular than it does on the tree model now some data sets; but the technology has not been optimized; therefore, necessary to provide a neural ...

AI Technical Summary

Problems solved by technology

[0002] In the data-driven era, deep learning algorithms have received extensive attention, and at the same time, machine learning is also booming. Deep learning is represented by deep neural network models, and machine learning is represented by tree models; although deep neural networks It has achieved great success in the fields of computer vision and natural language processing, but in many traditional areas, such as risk control models, anti-fraud and other tabular data-based tasks, the performance is not good; recently, Google proposed The neural network model TabNet, which has proved through experiments that it is indeed better than the popular tree model on some data sets; but it has not been technically optimized; therefore, it is necessary to propose a neural network algorithm for processing structured data, To at least partly solve the problems existing in the prior art

Images