Design And Implementation Of CNN Structured Feature Vector Input

Machine learning based prediction such as classification and regression,are ubiquitous in real-world applications.Despite the success,in general there is no single algorithm being a panacea across problems.Hence researchers and engineers have to carefully choose which algorithm is more suited to a given dataset.With the development of the times,deep learning technology has been developed for decades.Among them,Convolutional Neural Networks(CNNs)breaks through many various challenges in the field of computer vision,and also make many contributions to natural language processing and voice tasks.But CNNs are seldom used on traditional machine learning tasks where the feature dimension is not related to visual/audio/text data,where the data size is often smaller and its dimension is relatively low with no spatial/temporal dependency to each other.For example,in UCI machine learning repository,the size of data is usually from hundreds to thousands;at the same time,the original feature latitude is much smaller.Moreover,traditional data analysis tends to apply feature selection to find more compact and interpretable features.Such dimensionality reductions lose the opportunity to explore the complex interactions of features in a variety of non-linear ways.In order to make up for the above gap,this paper attempts to develop a convolutional neural network for vector data learning,even if the data itself has no spatial or temporal relationship.Given a one-dimensional feature vector of the input sample,we propose to transform the raw feature vector into multiple channels of feature maps.Specifically,the transformation involves padding,random permutation,reshaping,and concatenation,which finally leads to a few two-dimensional feature maps with multiple channels as standard inputs for 2-D ConveNets.We term such operations as a single shuffle adapter.In short,the main purpose of our research is to narrow the gap between convolution and traditional shallow learning models in learning medium-sized vector structure data,although traditional shallow models are still dominant.To the best of our knowledge,this is a new attempt to learn vector data through convolutional network networks,contrary to the fact that CNNs are primarily used for unstructured data(such as video,images),that is,spatial and temporal dependencies are ubiquitous.To this end,the ShuffleAdapter was designed to enrich the raw data while converting the vector into a friendly property map.We are then ready to propose and design a genetic searchbased approach to automatically discover the appropriate network structure so that the model can match a given data set.Finally,the experimental results also show that the convolutional neural network obtained by the above method is very good on structured data.