A review of addressing class noise problems of remote sensing classification

The development of image classification is one of the most important research topics in remote sensing. The prediction accuracy depends not only on the appropriate choice of the machine learning method but also on the quality of the training datasets. However, real-world data is not perfect and often suffers from noise. This paper gives an overview of noise filtering methods. Firstly, the types of noise and the consequences of class noise on machine learning are presented. Secondly, class noise handling methods at both the data level and the algorithm level are introduced. Then ensemble-based class noise handling methods including class noise removal, correction, and noise robust ensemble learners are presented. Finally, a summary of existing data-cleaning techniques is given.

A label noise filtering and label missing supplement framework based on game theory

Labeled data is widely used in various classification tasks.However,there is a huge challenge that labels are often added artificially.Wrong labels added by malicious users will affect the training effect of the model.The unreliability of labeled data has hindered the research.In order to solve the above problems,we propose a framework of Label Noise Filtering and Missing Label Supplement(LNFS).And we take location labels in Location-Based Social Networks(LBSN)as an example to implement our framework.For the problem of label noise filtering,we first use FastText to transform the restaurant's labels into vectors,and then based on the assumption that the label most similar to all other labels in the location is most representative.We use cosine similarity to judge and select the most representative label.For the problem of label missing,we use simple common word similarity to judge the similarity of users'comments,and then use the label of the similar restaurant to supplement the missing labels.To optimize the performance of the model,we introduce game theory into our model to simulate the game between the malicious users and the model to improve the reliability of the model.Finally,a case study is given to illustrate the effectiveness and reliability of LNFS.

Handling Label Noise in Air Traffic Complexity Evaluation Based on Confident Learning and XGBoost

Air traffic complexity is a critical indicator for air traffic operation,and plays an important role in air traffic management(ATM),such as airspace reconfiguration,air traffic flow management and allocation of air traffic controllers(ATCos).Recently,many machine learning techniques have been used to evaluate air traffic complexity by constructing a mapping from complexity related factors to air traffic complexity labels.However,the low quality of complexity labels,which is named as label noise,has often been neglected and caused unsatisfactory performance in air traffic complexity evaluation.This paper aims at label noise in air traffic complexity samples,and proposes a confident learning and XGBoost-based approach to evaluate air traffic complexity under label noise.The confident learning process is applied to filter out noisy samples with various label probability distributions,and XGBoost is used to train a robust and high-performance air traffic complexity evaluation model on the different label noise filtered ratio datasets.Experiments are carried out on a real dataset from the Guangzhou airspace sector in China,and the results prove that the appropriate label noise removal strategy and XGBoost algorithm can effectively mitigate the label noise problem and achieve better performance in air traffic complexity evaluation.

Label-noise robust classification with multi-view learning

Label noise is often contained in the training data due to various human factors or measurement errors,which significantly causes a negative effect on classifiers.Despite many previous methods that have been proposed to learn robust classifiers,they are mainly based on the single-view feature.On the other hand,although existing multi-view classification methods benefit from the more comprehensive information,they rarely consider label noise.In this paper,we propose a novel label-noise robust classification model with multi-view learning to overcome these limitations.In the proposed model,not only the classifier learning but also the label-noise removal can benefit from the multi-view information.Specifically,we relax the label matrix of the basic multi-view least squares regression model,and develop a nonlinear transformation with a natural probabilistic approximation in the process of labels,which is conveniently optimized and beneficial to improve the discriminative ability of classifiers.Moreover,we preserve the intrinsic manifold structure of multi-view data on the relaxed label matrix,facilitating the process of label relaxation.For optimizing the proposed model with the nonlinear transformation,we derive a lemma about the partial derivation of the softmax related function,and develop an efficient alternating algorithm.Experimental evaluations on six real-world datasets confirm the advantages of the proposed method,compared to the related state-of-the-art methods.

A robust optimization method for label noisy datasets based on adaptive threshold: Adaptive-k

The use of all samples in the optimization process does not produce robust results in datasets with label noise.Because the gradients calculated according to the losses of the noisy samples cause the optimization process to go in the wrong direction.In this paper,we recommend using samples with loss less than a threshold determined during the optimization,instead of using all samples in the mini-batch.Our proposed method,Adaptive-k,aims to exclude label noise samples from the optimization process and make the process robust.On noisy datasets,we found that using a threshold-based approach,such as Adaptive-k,produces better results than using all samples or a fixed number of low-loss samples in the mini-batch.On the basis of our theoretical analysis and experimental results,we show that the Adaptive-k method is closest to the performance of the Oracle,in which noisy samples are entirely removed from the dataset.Adaptive-k is a simple but effective method.It does not require prior knowledge of the noise ratio of the dataset,does not require additional model training,and does not increase training time significantly.In the experiments,we also show that Adaptive-k is compatible with different optimizers such as SGD,SGDM,and Adam.The code for Adaptive-k is available at GitHub.