Research On New Methods Of Transfer Learning And Collaborative Learning

In many machine learning tasks,the form of data is undergoing the transformation from single-source data to multi-source data.This transformation makes the traditional models no longer applicable,so it is necessary to develop new models.The two main multi-source data forms are multi-distributed multi-source data and multi-feature-set multi-source data,the corresponding data fusion strategies are transfer learning and collaborative learning.This paper proposes three new multi-source data fusion algorithms.Compared with the exsiting models for single-source and multi-source data,the proposed methods have a significant improvement in theory and performance.The main works of this paper are as follows:1)The first work is to develop a clustering method for multi-feature-set multi-source data.The existing collaborative clustering methdos are basically the improvement of prototype clustering,spectral clustering and other algorithms which are suitable for the data with measurable distance between samples.For co-occurrence data,the values of data represent the co-occurrence frequencies between samples and features,so the data of each source can be regarded as a joint probability distribution.Based on this assumption,the paper proposes an information-theoretic collaborative clustering method,which formalizes the problem into the form of information theory.Therefore,the method can uses joint probability distributions between samples and features instead of the distance information.The experimental results show the effectiveness of the proposed method for multi-source co-occurrence data.2)The second work is to develop a representation learning method for multi-distributed multi-source data.The core idea of this kind of methods is to match the distributions of multi-source data,and the core issue is the selection of feature transformation method.The existing methods are most based on kernel methods,whose defects is the lack of interpretability and the difficulty to be selected.In order to solve the problems,this paper proposes a tranfer representation learning method based on the fuzzy system.The proposed method takes the fuzzy system as a feature learning method.On the one hand,it makes the process of feature transformation more interpretable,on the other hand,it avoids the selection of kernel functions.The experiments verify the advantages of the proposed method compared with the existing methods in terms of interpretability and transfer performance.3)The third work is for a new multi-source data form which is rarely concerned in the previous research.The most important characteristic of multi-feature-set multi-source data is the existance of paired sampels,which is different from the multi-distributed multi-source data.In practical applications,there is a kind of multi-source data which combines these two characteristics,that is,there exist paired samples among multi-distributed multi-source data.In this paper,we propose a new data fusion method which combines the transfer leanring and collaborative learning.On this kind of data,the experiments show the significant advantages of the proposed method over the existing algorithms.