Font Size: a A A

Research On Machine Learning Methods In Environments With Changing Features

Posted on:2021-04-13Degree:DoctorType:Dissertation
Country:ChinaCandidate:B J HouFull Text:PDF
GTID:1368330647450644Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
Traditional machine learning usually assumes that the features of data of the learn-ing environment are fixed.However,in real applications,the features of data of the learning environment are likely to change with the change of the learning environment and traditional machine learning methods cannot deal with this situation effectively.This paper consists of the researches on machine learning methods in environments with changing features.It proposes and analyzes several related scenarios,and solves problems in these scenarios from the aspects of algorithm,theory and experiment.1.An online learning method in environment with synchronous changing fea-tures,FESL.This paper analyzes a basic problem of feature changing learning with streaming data – features change synchronously.To solve this problem,this paper proposes an online learning method named FESL.This method leverages the overlapping period to recover the vanishing features,and thus can utilize the recov-ered data to improve the model's performance on the new feature space.Both the theories and experiments validate the effectiveness of our proposed method.2.An online learning method in environment with asynchronous changing fea-tures,PUFE.This paper analyzes a more realistic problem of feature changing when learning with streaming data – features change asynchronously.To solve this problem,this paper proposes an online learning method named PUFE.This method transforms the situation with asynchronous feature changing into the syn-chronous one by matrix completion.This paper also gives the upper bound of the data quantity to be observed for matrix completion,optimizes the upper bound of the synchronous case,and verifies the effectiveness of PUFE through experiments.3.An online learning method in environment with changing features and limited resources,SF~2 EL.In view of the difficulty in obtaining label information under the feature changing situation,this paper proposes SF~2 EL method.This method makes use of the similarity between samples to assist learning,but this brings the problem of high storage cost.Given the limited storage resources of the device and the fact that different devices have different storage budgets,SF~2 EL can also learn adaptively to fit different storage budgets in the feature changing scenario.The validity of SF~2 EL method is verified by theories and experiments.4.A study on interpretability of RNN in environment with changing features,LISOR.We focus on RNN and investigate its interpretability under the feature changing scenario.This paper proposes LISOR to study the interpretability of RNN by transforming it into a more interpretable model,i.e.,Finite State Automata(FSA).The influence of different kinds and lengths features on FSA is also studied.Experimental results show that long features do not lead to better FSA,and there-fore do not lead to better interpretability.This is consistent with Occam's razor principle,which is that if it works,the simpler the better.
Keywords/Search Tags:machine learning, feature changing, streaming data, supervised learning, semi-supervised learning, storage fit, recurrent neural network(RNN), interpretability
PDF Full Text Request
Related items