Research On Online Transfer Learning Algorithm And Its Application On Text Classification

With the advent of the big data era and the widespread application of deep learning algorithms in industry,the construction of machine learning models began to change from theory-driven to data-driven.But when dealing with online problems,a large number of samples can't be used at the same time and the samples can't be saved,so the conventional offline algorithms are difficult to play a role.Online transfer learning algorithms can reduce the target domain online task training samples requirements by transferring the source domain knowledge,but also can assist the construction of the target domain online model.So online transfer learning algorithms have received a lot of attention.The goal of this thesis is to improve the existing framework of supervised online transfer learning algorithms applied on online text classification problem and to propose a new method for semisupervised online transfer learning problem.Most of the existing online transfer learning algorithm frameworks are constructed based on the multi-model weighted integration.However,the existing weight calculation methods will lead to model performance and weight mismatch.Aiming at the problems in the existing weight calculation methods,this thesis proposes a weight update formula based on sample weighting,which reduces the difference between source domain and target domain models in prior knowledge,and effectively alleviates the mismatch between model performance and weight in training.Aiming at the semi-supervised online transfer learning problem,an algorithm framework based on pseudo label method is proposed.The source domain models are used to generate pseudo labels for target domain unlabeled samples,and an improved Co-training mechanism is introduced to improve the reliability of pseudo labels,thus making full use of the target domain unlabeled samples.The main research result of this thesis is to put forward the improved supervised online transfer learning algorithm framework,to propose the solution to the problem of semisupervised online transfer learning,and to implement the algorithm frameworks specifically for the text classification task.The validity of the proposed algorithms is proved by comparing experiments on authoritative data sets in the field of transfer learning.