One common problem with geographic data is that, for a specific geographic event, only occurrence information is available; information about the absence of the event is not available. We refer to ...these specific types of geospatial data as geographic one-class data (GOCD). Predicting the potential spatial distributions that a particular geographic event may occur from GOCD is difficult because traditional binary classification methods that require availability of both positive and negative training samples cannot be used. The objective of this research is to define GOCD and propose novel approaches for modelling potential spatial distributions of geographic events using GOCD. We investigate the effectiveness of one-class support vector machine (OCSVM), maximum entropy (MAXENT) and the newly proposed positive and unlabelled learning (PUL) algorithm for solving GOCD problems using a case study: species distribution modelling from synthetic data. Our experimental results indicate that generally OCSVM, MAXENT and PUL are effective in modelling the GOCD. Each method has advantages and disadvantages, but PUL seems to be the most promising method.
Positive and unlabeled (PU) learning is aimed at building a binary classifier to distinguish target from background using only the known positive samples, which is an advanced solution for the ...hyperspectral target detection task. However, when PU learning meets complex hyperspectral scenarios, there are two main challenges: 1) How to estimate the class prior accurately. The class prior, i.e., the target proportion, is an important prior for PU learning to learn the discriminant boundary, but it is difficult to estimate in hyperspectral imagery, due to the inter-class spectral similarity. 2) How to remove redundancy and improve the discriminative features of the target. The diagnostic spectral feature extraction is important for the weakly supervised PU learning models as it can help with separating target from background. In this paper, to tackle these challenges, a robust PU learning framework with key band selection (PU-KBS) is proposed, which is modeled as an end-to-end and class prior free PU learning framework, where the accurate class prior and the most discriminative key band subset are jointly initialized and iteratively updated until reaching the optimal result by evolutionary search. Meanwhile, a deep PU learning detector is introduced for guiding the subsequent search direction and discriminative deep feature extraction. The proposed PU-KBS framework was verified using different hyperspectral datasets, where accurate class prior estimation, diagnostic spectral characteristics, and robust detection results could be obtained simultaneously by the PU-KBS framework. Furthermore, the improvement in band selection interpretability and detection performance was proven experimentally.
•High-resolution annual cropland maps were made for the Huang-Huai-Hai Plain.•A prior information was used to guide image composition showing superior performance.•A new method was proposed for ...sample migration cross years.•Mapping with Random Forest and Positive and Unlabeled Learning methods combination.
Numerous map products containing cropland data are currently available. However, due to their low temporal and spatial resolutions and limited accuracy, their practical applications are limited. Thus, research on the precise mapping of croplands with high temporal and spatial resolutions is still urgently needed. The Huang-Huai-Hai Plain, a critical grain production area in China, was used as the research area to produce high-resolution annual cropland maps from 2018 to 2022. Sentinel-2 imagery, digital elevation model (DEM) data, and China Agricultural Ecological Zoning (CAEZ) data were collected. Based on these data, the study area was stratified into different zones, and then the distinct characteristics and optimal timing for differentiating cropland from other land cover types within each zone were analyzed and used as prior information for guiding time-series image compositing. Furthermore, the spectral information divergence (SID) index was used to modify the existing sample migration technique to transfer training samples across years. Via the Google Earth Engine (GEE) platform, the random forest (RF) classification method and positive and unlabeled learning (PUL) method were combined to construct cropland maps from composited images and terrain information from DEMs. Compared with the seasonal or spectral-temporal metric (STM) image composition methods, the image composition method based on prior information significantly improved the distinction between cropland and other land cover types. Moreover, the newly proposed sample migration strategy substantially increased the accuracy of training samples transferred across years. Compared with existing cropland maps from products such as the ESA, ESRI, GlobeLand30, GLAD, and CLCD, the cropland maps generated in this study demonstrated superior accuracy, with average overall accuracy (OA) values from 88.81% to 91.66% for different zones, thus outperforming other products with OA values from 77.53% to 88.90%. This study supports the development of related agricultural and ecological research and provides a valuable reference for cropland mapping.
•A deep one-class crop classification framework is proposed for multi-modal imagery.•DOCC takes one target crop samples as the input thus avoiding redundant labeling.•A one-class crop extraction loss ...is designed for positive and unlabeled learning.•Experiment reveals the optimal data source to identify winter wheat and rapeseed.
Large-scale crop mapping is an important task in agricultural resource monitoring, but it does usually require the ground-truth labels of all the land-cover types in the remotely sensed imagery. However, labeling each land-cover type is time-consuming and labor-intensive. One-class classification, which only needs samples of the class of interest, can solve the problem of redundant labeling. However, the traditional one-class classifiers require well-designed features to realize fine classification, and are thus difficult to apply to complex multi-modal remote sensing data, i.e., optical imagery and synthetic aperture radar (SAR) imagery. In this paper, a deep one-class crop (DOCC) framework that includes a deep one-class crop extraction module and a one-class crop extraction loss module is proposed for large-scale one-class crop mapping. The DOCC framework takes only the samples of one target class as the input to extract the crop of interest by positive and unlabeled learning and can automatically extract the features for one-class crop mapping, without requiring a large amount of labeling for all the land-cover type or feature design based on prior expert knowledge. Experiments conducted on multi-modal remote sensing data, i.e., Zhuhai-1 hyperspectral satellite data, Sentinel-2 multispectral time-series satellite data and Sentinel-1 SAR satellite data, illustrate that DOCC can automatically extract the effective features for one-class classification from multi-modal satellite imagery and reaches the highest F1 accuracy compared with other methods on respective satellite imagery. The results also reveal the different performance of multi-modal satellite imagery when they are used to extract different crop types. Meanwhile, the feasibility of DOCC for multi-modal data can be beneficial for large-scale mapping under different conditions when the samples of multi-class are difficult to obtain.
•Proposal of a graph-based method for Positive and Unlabeled Learning that uses graph-based strategies in all steps.•Effectively use of unlabeled documents to improve classification ...performances.•Evaluation of 3 usual graph construction methods and 2 label propagation algorithms.•The proposal achieves better results with statistically significant differences (SSD) compared to state-of-art related approaches.•The proposal achieves better results with SSD than methods based on vector space model and graphs.
Positive and Unlabeled Learning (PUL) uses unlabeled documents and a few positive documents for retrieving a set of “interest” documents from a text collection. Usually, PUL approaches are based on the vector space model. However, when dealing with semi-supervised learning for text classification or information retrieval, graph-based approaches have been proved to outperform vector space model-based approaches. So, in this article, a graph-based approach for PUL is proposed: Label Propagation for Positive and Unlabeled Learning (LP-PUL). The proposed framework consists of three steps: (i) building a similarity graph, (ii) identifying reliable negative documents, and (iii) performing label propagation to classify the remaining unlabeled documents as positive or negative. We carried out experiments to measure the impact of the different choices in each step of the proposed framework. We also demonstrated that the proposal surpasses the classification performance of other PUL (RC-SVM, PU-LP, and PE-PUC) or one-class learning (k-NN-based, k-Means-based, and Dense Autoencoder) algorithms in terms of F1. Considering the best results of any algorithm used in the experimental evaluation, PU-PUL can improve the classification performance from 2%, when using only 1 labeled document, to 28%, when 30 labeled documents are employed.
Fake news detection (FND) tools are essential to increase the reliability of information in social media. FND can be approached as a machine learning classification problem so that discriminative ...features can be automatically extracted. However, this requires a large news set, which in turn implies a considerable amount of human experts' effort for labeling. In this paper, we explore Positive and Unlabeled Learning (PUL) to reduce the labeling cost. In particular, we improve PUL with the network-based Label Propagation (PU-LP) algorithm. PU-LP achieved competitive results in FND exploiting relations between news and terms and using few labeled fake news. We propose integrating an attention mechanism in PU-LP that can define which terms in the network are more relevant for detecting fake news. We use GNEE, a state-of-the-art algorithm based on graph attention networks. Our proposal outperforms state-of-the-art methods, improving F1 in 2% to 10%, especially when only 10% labeled fake news are available. It is competitive with the binary baseline, even when nearly half of the data is labeled. Discrimination ability is also visualized through t-SNE. We also present an analysis of the limitations of our approach according to the type of text found in each dataset.
•We propose a model-less ranking method when multiple representations of the data are available.•We study the method in principled simulation settings that could be amenable for future analysis in ...follow-up work.•We evaluate the method in realistic settings and demonstrate that it is preferred to the existing techniques that handle distance-based ranking.
Learning to rank – producing a ranked list of items specific to a query and with respect to a set of supervisory items – is a problem of general interest. The setting we consider is one in which no analytic description of what constitutes a good ranking is available. Instead, we have a collection of representations and supervisory information consisting of a (target item, interesting items set) pair. We demonstrate analytically, in simulation, and in real data examples that learning to rank via combining representations using an integer linear program is effective when the supervision is as light as “these few items are similar to your item of interest.” While this nomination task is quite general, for specificity we present our methodology from the perspective of vertex nomination in graphs. The methodology described herein is model agnostic.
Positive and Unlabeled learning (PU learning) aims to train a binary classifier solely based on positively labeled and unlabeled data when negatively labeled data are absent or distributed too ...diversely. However, none of the existing PU learning methods takes the class imbalance problem into account, which significantly neglects the minority class and is likely to generate a biased classifier. Therefore, this paper proposes a novel algorithm termed “Cost-Sensitive Positive and Unlabeled learning” (CSPU) which imposes different misclassification costs on different classes when conducting PU classification. Specifically, we assign distinct weights to the losses caused by false negative and false positive examples, and employ double hinge loss to build our CSPU algorithm under the framework of empirical risk minimization. Theoretically, we analyze the computational complexity, and also derive a generalization error bound of CSPU which guarantees the good performance of our algorithm on test data. Empirically, we compare CSPU with the state-of-the-art PU learning methods on synthetic dataset, OpenML benchmark datasets, and real-world datasets. The results clearly demonstrate the superiority of the proposed CSPU to other comparators in dealing with class imbalanced tasks.
We consider the problem of estimating the
class prior
in an unlabeled dataset. Under the assumption that an additional labeled dataset is available, the class prior can be estimated by fitting a ...mixture of class-wise data distributions to the unlabeled data distribution. However, in practice, such an additional labeled dataset is often not available. In this paper, we show that, with additional samples coming only from the positive class, the class prior of the unlabeled dataset can be estimated correctly. Our key idea is to use properly penalized divergences for model fitting to cancel the error caused by the absence of negative samples. We further show that the use of the penalized
L
1
-distance gives a computationally efficient algorithm with an analytic solution. The consistency, stability, and estimation error are theoretically analyzed. Finally, we experimentally demonstrate the usefulness of the proposed method.
Application of supervised machine learning algorithms for mineral prospectivity mapping (MPM) requires positive and negative training samples. Typically, known mineral deposits are considered as ...positive training samples. However, the selection of negative training samples in the process of MPM is challenging. The one-class classification methods require positive and unlabeled samples or only positive samples; while without requiring negative training samples. In this study, the positive and unlabeled learning (PUL) algorithm was employed to produce a potential map for Fe polymetallic mineralization in southwestern Fujian province, China. This study first examined the sensitivity of the PUL algorithm to different training sets of labeled and unlabeled locations. The predictive results on 10 random unlabeled datasets confirm that PUL modeling with different training sets is reproducible and stable. In addition, the trained model provides a strong spatial correlation between the predictive variables and the locations of known mineral deposits. Finally, the performance of PUL1algorithm is compared to one-class support vector machine (OCSVM), artificial neural networks (ANN), and logistic regression (LR). Comparative results indicate that the PUL model can achieve a better performance in terms of both fitting-rate, prediction-rate, and AUC value compared with OCSVM, ANN and LR. The labelling efforts can be significantly reduced because the PUL algorithm requires only a small number of positive samples and utilizes unlabeled data in training.
•The positive and unlabeled learning (PUL) algorithm was employed for the mapping of mineral prospectivity.•The PUL algorithm delivers a better potential map than the classical OCSVM and ANN for MPM.•A case study from southwestern Fujian district was conducted.
PDF
