Study On The Application Of Statistical Learning Methods To Triple Classification Problems

With the development of artificial intelligence technology,statistical learning methods have been maturely applied in many fields.As the core field of statistical learning methods,classification models have become a hot research topic today.Existing classification models are usually designed for binary classification problems.With the continuous refinement of application areas,new challenges have been raised for conventional classification models.Existing classification methods often have difficulty achieving satisfactory performance when dealing with multiclass labels,time series features,and high-dimensional sparse data.Therefore,proposing a classification method that can effectively handle high-dimensional large-scale data and can be extended to multi-class scenarios has high practical value.This thesis combines statistical learning methods with deep learning methods to propose an effective ENet-LSTM/DNN combination model for solving classification problems.ENet/SCAD/MCP penalty logit models,machine learning models,and deep learning models are introduced as comparative objects.The performance of the ENet-LSTM/DNN model in stock price prediction and disease classification problems is evaluated using Kappa coefficient,PDI index,HUM index,and ROC curve.In the stock price prediction problem,this thesis obtained historical stock price data and financial news data for three companies,AAPL,MNST and BAC,from the Nasdaq index.Based on the TTR package and text sentiment analysis,32 technical indicators and 2 market sentiment indicators were calculated as predictive variables,with the rise,plateau,and decline of stock prices as response variables.Eight prediction methods were compared for their predictive performance,and the results showed that the ENet-LSTM model had the best predictive performance,with a predictive accuracy of up to 0.6350.The ENet-LSTM model combines the advantages of the ENet penalty method and the LSTM framework,which can eliminate multicollinearity among data and accurately capture effective information in time series data using gate mechanisms.In the disease classification problem,this thesis selected the infant genetic disease dataset from the Kaggle database as the research object.The physiological indicators of patients were taken as predictive variables,and the genetic disease types of patients were taken as response variables.The ENet-DNN model was established,and seven classification models were introduced for comparison.The results showed that the predictive accuracy of the ENet-DNN model was up to 0.702,which could effectively handle high-dimensional large-scale datasets and had better predictive performance than single ENet penalty logit model and DNN model.Therefore,based on statistical learning methods and deep learning methods,this thesis proposed ENet-LSTM/DNN classification methods that combine the advantages of ENet penalty method and deep learning models.These methods achieved good predictive performance in three classification problems in stock price prediction and disease field,and have high research value and practical significance.