Towards The Design And Exploration Of Deep Neural Networks For Tabula Data

The neural network method originated in the last century and has undergone a lot of adjustments,improvements and optimizations in the long-term development process.In recent years,It finally has achieved remarkable results,and has gradually been favored by many scholars from different fields.With the consideration on the characteristics of data in different fields,a huge quantity of new neural network-based models have been proposed,and these new neural network methods gain the significant improvements in their respective fields.At the same time,the data processing capabilities of neural networks have also increased with the deepening of neural networks.Many common deep neural network architectures have been proposed and widely used as skeletons in many different tasks.On various complex human cognition tasks,the development of deep neural networks is changing rapidly,and has become the default research direction.But,at the same time,as a very common prediction task,the learning and prediction of tabulated data seems to be abandoned by deep learning methods,and is still in the dominance of traditional machine learning methods.The reason why deep learning is not applied to tabulated data may be roughly summarized into two points: On the one hand,tabulated data usually has a small amount of data samples and feature dimensions.For the neural networks which has strong fitting capability even over-fitting tendency,the small data size is not friendly to the training of neural network model;on the other hand,the features of a single sample of tabulated data are usually composed of fixed-length feature vector whose components come from the measurements of the target sample through different observation viewpoints,thus there is little spatial/temporal local correlation on tabulated data feature.This is also difficult for researchers to find the suitable neural network structures for tabulated datasets,which makes the network methods unable to achieve effective prediction accuracy.But at the same time,traditional machine learning methods are also not a panacea in this field.Traditional methods usually have a strong requirements on the quality of dataset,thus it take the feature engineering as an important pre-step before model training,and in the most cases,a specific traditional machine learning method only has good performance on a certain type of dataset.Therefore,applying traditional methods not only places high demands on researchers,but also adds a lot of extra overhead.The generality and scalability of neural networks provide potential possibilities for solving this problem.Based on the above analysis,this article mainly explores the application of neural network methods on the tabulated data.First of all,we discussed many traditional machine learning methods which has been the inspiration for our neural network designing,and then we examines many kinds of the neural network techniques.After that,we comprehensively analyzes the feasibility of different methods and techniques for tabulated data analysis and prediction.Secondly,based on the existing methods,we proposed a novel convolution neural network architecture which is suitable for tabulated data learning,and designed a variety of fully connected neural networks for experimental comparison.Third,consider that the tabulated data can vary greatly on the data size and feature length,we transformed the genetic algorithm and applied it to the adaptive search of the structural parameters of our proposed neural network methods.Finally,we exam the traditional machine learning method and our neural network methods on various kinds of tabulated datasets.The effectiveness of our convolution neural network method is verified through extensive experiments.