Applications Research Of Multi-view And Transfer Learning Based Classification Methods

Since the modern era is in the period of information explosion,machine learning has also achieved significant development,especially in the fields of clustering,classification and regression.Classification is still one of the most important research fields in today's machine learning,and it plays an important role in practical applications such as semantic analysis,image recognition,and face recognition.However,with the continuous development and advancement of society and the innovation in the computer field,the amount of data continues to increase.The application scenarios of many emerging data are continuously expanded.This brings new problems which many traditional machine learning methods cannot adapt: the collected data is in a variety of forms,and its internal relationship structure is complex;the amount of collected data with lables is very limited.These may cause the classic classification methods to be noneffective when applied to the application scenarios.Then,these methods would meet more severe challenges: on one hand,in terms of the diversity of data,the existing classical classification method ignores the intrinsic relationship of various forms of data,which leads to the inability of such methods to obtain satisfactory results when processing such data.On the other hand,due to the serious lack of the collectable labeled samples,the traditional classification method can not obtain better generalization ability when the data is directly processed and modeled,and the effect is poor.Therefore,this study mainly solves the problems of how to extend the classic machine learning classification algorithm under the challenge of emerging data,in order to cope with the above problems in the emerging data application scenarios,and finally get better classification performance.In order to solve the above challenges that the classical classification method needs to address in the emerging data application scenarios,this study focuses on two application scenarios: multi-view learning and transfer learning,and responds to the expansion and improvement of traditional classification methods,so as to meet the chanllenges of emerging data.The details of this study are as follows:(1)The first part is the Section 2 and Section 3.It mainly studies the classification method based on multi-view learning and its application.Firstly,in Section 2,a novel multi-view support vector machine based on consistent hidden density distributions between views in common hidden space is proposed to deal with the challenge that the recognition effect for multi-view face recognition is poor due to the difference between the samples from different views,and its theoretical convergence is also analyzed.The proposed method does not directly calculate the difference between samples from different views,but base on the probability density consistency in hidden space and use support vector machine to solve the problem cleverly.In addition,a new optimization method based on the bundle method is used to optimize the optimization problem effectively and quickly.The proposed method is successfully evaluated for multi-view face recognition with good practicability and validity,especially in special multi-view scenarios where the dissimilarity between the samples from the same class but different views is greater than that between the samples from the different classes of same view.Then,in Section 3,a novel multi-view local linear k nearest neighbor method is proposed and the theoretical and experimental analysis of the method are also presented.MV-LLKNN is developed under two reliable and theoretically provable assumptions without any explicit prediction model for multi-view learning.The proposed method exploits the discriminative nature of sparse to perform classification.MV-LLKNN can realize its effective prediction for a multi-view test sample by cheaply using both on-hand FISTA and KNN.In addition,it has been proved that the proposed method has other interesting properties,such as CENN which can enhance the correlation between samples in same class and connection of Bayesian decision rules.(2)The Section 4 and Section 5 is the second part.It mainly studies the classification method based on transfer learning and its application.In Section 4,aiming at the serious shortage of training data in the current scene where the classification performance is degraded,a DLSR-based inductive transfer learning algorithm is proposed for multi-class classification tasks.The algorithm is based on the framework of the inductive transfer learning algorithm.However,it is different from other inductive transfer learning algorithms that directly utilize the data or features of the source domain,but utilizes a knowledge leverage mechanism that transfer knowledge from the source domain.Some kind of knowledge is used for the target domain,and this can well protect the security of the source domain data,and can simultaneously utilize the data of the target domain and the knowledge from the source domain,so that the performance is better.The algorithm inherits the characteristics of DLSR by enlarging the margin between different classes to better adapt to the characteristics of multi-class classification scenarios,and expands it into a new method with certain transfer learning ability,and transfers knowledge from the source domain.These innovations ensure the learned model can be more reasonable and effective.Then,in Section 5,a novel inductive transfer learning method based on both knowledge and label space transfer is proposed for multiclass epileptic EEG signal recognition.It not only uses the proposed LSR based classifier ?-LSR to enlarge the label margins between different classes in the corresponding label space so as to classify multiclass samples effectively,but also leverages both knowledge and label space information of source domain by means of inductive transfer learning to lift up the classification performance of insufficient target data.Since this method is designed on a linear model,the resultant prediction function can reveal/explain the feature relevance clearly with a strong theoretical guarantee.Extensive experimental studies indicate the power of the proposed method for the recognition of insufficient multiclass epileptic EEG signals.