Search

Handwriting difficulty is a defining feature of Chinese developmental dyslexia (DD) due to the complex structure and dense information contained within compound characters. Despite previous attempts to use deep neural network models to extract handwriting features, the temporal property of writing characters in sequential order during dictation tasks has been neglected. By combining transfer learning of convolutional neural network (CNN) and positional encoding with the temporal-sequential encoding of long short-term memory (LSTM) and attention mechanism, we trained and tested the model with handwriting images of 100,000 Chinese characters from 1064 children in Grades 2–6 (DD = 483; Typically Developing [TD] = 581). Using handwriting features only, the best model reached 83.2% accuracy, 79.2% sensitivity, 86.4% specificity, and 91.2% AUC. With grade information, the best model achieved 85.0% classification accuracy, 83.3% sensitivity, 86.4% specificity, and 89.7% AUC. These findings suggest the potential of utilizing machine learning technology to identify children at risk for dyslexia at an early age. [ABSTRACT FROM AUTHOR]

Malicious domains provide malware with covert communication channels which poses a severe threat to cybersecurity. Despite the continuous progress in detecting malicious domains with various machine learning algorithms, maintaining up‐to‐date various samples with fine‐labeled data for training is difficult. To handle these issues and improve the detection accuracy, a novel malicious domain detection method named MDND‐SS‐PO is proposed that combines semi‐supervised learning and parameter optimization. The contributions of the study are as follows. First, the method extracts the statistical features of the IP address, TTL value, the NXDomain record, and the domain name query characteristics to discriminate Domain‐Flux and Fast‐Flux domain names simultaneously. Second, an improved DBSCAN based on the neighborhood division is designed to cluster labeled data and unlabeled data with low time consumption. Then, based on the clustering hypothesis, unlabeled data is tagged with pseudo‐label according to the cluster results, which aims to train a supervised classifier effectively. Finally, Gaussian process regression is used to optimize parameter settings of the algorithm. And the Silhouette index and F1 score are introduced to evaluate the optimization results. Experimental results show that the proposed method achieved a precise detection performance of 0.885 when the ratio of labeled data is 5%. [ABSTRACT FROM AUTHOR]

Seasonal freezing depth change is important in many environmental science and engineering applications. However, such changes are rare at region scales, especially over China, in the long time series. In this study, we evaluated the annual changes in seasonal maximum freezing depth (MFD) over China from 1971 to 2020 using an ensemble-modeling method based on support vector machine regression (SVMR) with 600 repetitions. Remote sensing data and climate data were input variables used as predictors. The models were trained using a large amount of annual measurement data from 600 meteorological stations. The main reason for using SVMR here was because it has been shown to perform better than random forests (RF), k-nearest neighbors (KNN), and generalized linear regression (GLR) in these cases. The prediction results were generally consistent with the observed MFD values. Cross validation showed that the model performed well on training data and had a better spatial generalization ability. The results show that the freezing depth of seasonally frozen ground in China decreased year by year. The average MFD was reduced by 3.64 cm, 7.59 cm, 5.54 cm, and 5.58 cm, in the 1980s, 1990s, 2000s, and 2010s, respectively, compared with the decade before. In the last 50 years, the area occupied by the freezing depth ranges of 0–40 cm, 40–60 cm, 60–80 cm, 80–100 cm, and 120–140 cm increased by 99,300 square kilometers, 146,200 square kilometers, 130,300 square kilometers, 115,600 square kilometers, and 83,800 square kilometers, respectively. In addition to the slight decrease in freezing depth range of 100–120 cm, the reduced area was 29,500 square kilometers. Freezing depth ranges greater than 140 cm showed a decreasing trend. The freezing depth range of 140–160 cm, which was the lowest range, decreased by 89,700 square kilometers. The 160–180 cm range decreased by 120,500 square kilometers, and the 180–200 cm range decreased by 161,500 square kilometers. The freezing depth range greater than 200 cm, which was the highest reduction range, decreased by 174,000 square kilometers. Considering the lack of data on the change in MFD of seasonally frozen ground in China in recent decades, machine learning provides an effective method for studying meteorological data and reanalysis data in order to predict the changes in MFD. [ABSTRACT FROM AUTHOR]

Attention mechanisms have significantly boosted the performance of video classification neural networks thanks to the utilization of perspective contexts. However, the current research on video attention generally focuses on adopting a specific aspect of contexts (e.g., channel, spatial/temporal, or global context) to refine the features and neglects their underlying correlation when computing attentions. This leads to incomplete context utilization and hence bears the weakness of limited performance improvement. To tackle the problem, this paper proposes an efficient attention-in-attention (AIA) method for element-wise feature refinement, which investigates the feasibility of inserting the channel context into the spatio-temporal attention learning module, referred to as CinST, and also its reverse variant, referred to as STinC. Specifically, we instantiate the video feature contexts as dynamics aggregated along a specific axis with global average and max pooling operations. The workflow of an AIA module is that the first attention block uses one kind of context information to guide the gating weights calculation of the second attention that targets at the other context. Moreover, all the computational operations in attention units act on the pooled dimension, which results in quite few computational cost increase (< 0.02%). To verify our method, we densely integrate it into two classical video network backbones and conduct extensive experiments on several standard video classification benchmarks. The source code of our AIA is available at https://github.com/haoyanbin918/Attention-in-Attention. [ABSTRACT FROM AUTHOR]

Previous deep learning methods usually required large‐scale annotated data, which is computationally exhaustive and unrealistic in certain scenarios. Therefore, few‐shot learning, where only a few annotated training images are available for training, has attracted increasing attention these days, showing huge potential in practical applications, such as portable equipment or security inspection, and so on. However, current few‐shot learning methods usually neglect the valuable semantic correlations between samples, thereby failing in extracting discriminating relations to achieve accurate predictive results. In this work, extending on a recent state‐of‐the‐art few‐shot learning method, transductive relation‐propagation network (TRPN), which considers the correlations between training samples, a constrained relation‐propagation network is proposed to further regularise the distilled correlations and thus achieve favourable few‐shot classification performance. The proposed framework contains three main components, namely preprocess module, relational propagation module, and relation constraint module. First, sample features are extracted and a relation graph node is constructed by treating the relation of each support–query pair as a graph node in the preprocess module. After that, in the relation propagation module (RPM), the valuable information of support–query pairs is modelled and propagated to directly generate the relational representations for further prediction. Then, a relation constraint module is introduced to regularise the relational representations and make it consistent with the ground‐truth relations as much as possible. With the guidance of the effective RPM and relation constraint module, the relational representations of the support–query pairs are distinguishable and thus can achieve accurate predictive results. Comprehensive experiments conducted on widely used benchmarks validate the effectiveness of our method compared to state‐of‐the‐art few‐shot classification approaches. [ABSTRACT FROM AUTHOR]

Recent studies in brain science and neurological medicine paid a particular attention to develop machine learning-based techniques for the detection and prediction of epileptic seizures with electroencephalogram (EEG). As a noninvasive monitoring method to record brain electrical activities, EEG has been widely used for capturing the underlying dynamics of disruptive neuronal responses across the brain in real-time to provide clinical guidance in support of epileptic seizure treatments in practice. In this study, we introduce a novel dynamic learning method that first infers a time-varying network constituted by multivariate EEG signals, which represents the overall dynamics of the brain network, and subsequently quantifies its topological property using graph theory. We demonstrate the efficacy of our learning method to detect relatively strong synchronization (characterized by the algebraic connectivity metric) caused by abnormal neuronal firing during a seizure onset. The computational results for a realistic scalp EEG database show a detection rate of 93.6% and a false positive rate of 0.16 per hour (FP/h); furthermore, our method observes potential pre-seizure phenomena in some cases. [ABSTRACT FROM AUTHOR]

Artificial intelligence (AI) tools have been applied to diagnose or predict disease risk from medical images with recent data disclosure actions, but few of them are designed for mobile terminals due to the limited computational power and storage capacity of mobile devices. In this work, a novel AI diagnostic system is proposed for cholelithiasis recognition on mobile devices with Android platform. To this aim, a data set of CT images of cholelithiasis is firstly collected from The Third Hospital of Shandong Province, China, and then we technically use histogram equalization to preprocess these CT images. As results, a lightweight convolutional neural network is obtained in a constructive way to extract cholelith features and recognize gallstones. In terms of implementation, we compile Java and C++ to adapt to the application of deep learning algorithm on mobile devices with Android platform. Noted that, the training task is completed offline on PC, but cholelithiasis recognition tasks are performed on mobile terminals. We evaluate and compare the performance of our MobileNetV2 with MobileNetV1, Single Shot Detector (SSD), YOLOv2 and original SSD (with VGG-16) as feature extractors for object detection. It is achieved that our MobileNetV2 achieve similar accuracy rate, about 91% with the other four methods, but the number of parameters used is reduced from 36.1M (SSD 300, SSD512), 50.7M (Yolov2) and 5.1M (MobileNetV1) to 4.3M (MobileNetV2). The complete process on testing mobile devices, including Virtual machine, Xiaomi 7 and Htc One M8 can be controlled within 4 seconds in recognizing cholelithiasis as well as the degree of the disease. [ABSTRACT FROM AUTHOR]

• Computing on edge end improves the efficiency and reliability of data analysis. • A lightweight CNN model is efficient for real-time crowd density estimation on edge. • Better crowd density inference speed with a slight increase in estimation accuracy. • Equip the model in an IoT device to monitor the crowd density in a Subway Station. Crowd stampedes and incidents are critical threats to public security that have caused countless deaths during the past few decades. To avoid crowd stampedes, real-time crowd density estimation can help monitor crowd movements, and thus support a timely evacuation strategy development. In previous studies, scholars and engineers developed multiple video-based crowd density estimation algorithms based on deep neural networks. The excessive computational complexity of deep learning algorithms exacerbated the algorithm's efficiency, causing unacceptable real-time performance. In the Internet of Things era, deploying the crowd density estimation task with edge computing is an advanced strategy to maintain the real-time performance of the entire system. Considering the limited computational resources on the edge devices, deep learning-based crowd density estimation algorithms normally cannot be handled. To fulfill the deployment on the edge device, the algorithms need to be optimized with a smaller model size. Therefore, this paper proposes a lightweight Convolutional Neural Networks (CNN) based crowd density estimation model by combining the modified MobileNetv2 and the dilated convolution. Public crowd image data sets are used to conduct experiments for evaluating the performance of the proposed algorithm in terms of accuracy and inference speed. The results show that our model achieves much better inference speed accompanied by a slight increase in accuracy. The proposed method of this study can enhance the performance of the crowd monitoring system, and therefore help avoid crowd stampedes and incidents. [ABSTRACT FROM AUTHOR]

Online class imbalance learning is a new learning problem that combines the challenges of both online learning and class imbalance learning. It deals with data streams having very skewed class distributions. This type of problems commonly exists in real-world applications, such as fault diagnosis of real-time control monitoring systems and intrusion detection in computer networks. In our earlier work, we defined class imbalance online, and proposed two learning algorithms OOB and UOB that build an ensemble model overcoming class imbalance in real time through resampling and time-decayed metrics. In this paper, we further improve the resampling strategy inside OOB and UOB, and look into their performance in both static and dynamic data streams. We give the first comprehensive analysis of class imbalance in data streams, in terms of data distributions, imbalance rates and changes in class imbalance status. We find that UOB is better at recognizing minority-class examples in static data streams, and OOB is more robust against dynamic changes in class imbalance status. The data distribution is a major factor affecting their performance. Based on the insight gained, we then propose two new ensemble methods that maintain both OOB and UOB with adaptive weights for final predictions, called WEOB1 and WEOB2. They are shown to possess the strength of OOB and UOB with good accuracy and robustness. [ABSTRACT FROM AUTHOR]

To facilitate software testing, and save testing costs, a wide range of machine learning methods have been studied to predict defects in software modules. Unfortunately, the imbalanced nature of this type of data increases the learning difficulty of such a task. Class imbalance learning specializes in tackling classification problems with imbalanced distributions, which could be helpful for defect prediction, but has not been investigated in depth so far. In this paper, we study the issue of if and how class imbalance learning methods can benefit software defect prediction with the aim of finding better solutions. We investigate different types of class imbalance learning methods, including resampling techniques, threshold moving, and ensemble algorithms. Among those methods we studied, AdaBoost.NC shows the best overall performance in terms of the measures including balance, G-mean, and Area Under the Curve (AUC). To further improve the performance of the algorithm, and facilitate its use in software defect prediction, we propose a dynamic version of AdaBoost.NC, which adjusts its parameter automatically during training. Without the need to pre-define any parameters, it is shown to be more effective and efficient than the original AdaBoost.NC. [ABSTRACT FROM AUTHOR]

Background: Acute pancreatitis (AP) with critical illness is linked to increased morbidity and mortality. Current risk scores to identify high-risk AP patients have certain limitations.Objective: To develop and validate a machine learning tool within 48 h after admission for predicting which patients with AP will develop critical illness based on ubiquitously available clinical, laboratory, and radiologic variables.Methods: 5460 AP patients were enrolled. Clinical, laboratory, and imaging variables were collected within 48 h after hospital admission. Least Absolute Shrinkage Selection Operator with bootstrap method was employed to select the most informative variables. Five different machine learning models were constructed to predictive likelihood of critical illness, and the optimal model (APCU) was selected. External cohort was used to validate APCU. APCU and other risk scores were compared using multivariate analysis. Models were evaluated by area under the curve (AUC). The decision curve analysis was employed to evaluate the standardized net benefit.Results: Xgboost was constructed and selected as APCU, involving age, comorbid disease, mental status, pulmonary infiltrates, procalcitonin (PCT), neutrophil percentage (Neu%), ALT/AST, ratio of albumin and globulin, cholinesterase, Urea, Glu, AST and serum total cholesterol. The APCU performed excellently in discriminating AP risk in internal cohort (AUC = 0.95) and external cohort (AUC = 0.873). The APCU was significant for biliogenic AP (OR = 4.25 [2.08-8.72], P < 0.001), alcoholic AP (OR = 3.60 [1.67-7.72], P = 0.001), hyperlipidemic AP (OR = 2.63 [1.28-5.37], P = 0.008) and tumor AP (OR = 4.57 [2.14-9.72], P < 0.001). APCU yielded the highest clinical net benefit, comparatively.Conclusion: Machine learning tool based on ubiquitously available clinical variables accurately predicts the development of AP, optimizing the management of AP. [ABSTRACT FROM AUTHOR]

The use of multiparametric magnetic resonance imaging (mpMRI)-derived radiomics has the potential to offer noninvasive, imaging-based biomarkers for the identification of subvisual characteristics indicative of a poor oncologic outcome. The present study, therefore, seeks to develop, validate, and assess the performance of an MRI-derived radiomic model for the prediction of prostate cancer (PC) recurrence following radical prostatectomy (RP) with curative intent. mpMRI imaging was obtained from 251 patients who had undergone an RP for the treatment of localized prostate cancer across two institutions and three surgeons. All patients had a minimum of 2 years follow-up via prostate-specific antigen serum testing. Each prostate mpMRI was individually reviewed, and the prostate was delineated as a single slice (ROI) on axial T2 high-resolution image sets. A total of 924 radiomic features were extracted and tested for stability via intraclass correlation coefficient (ICC) following image normalization via histogram matching. Fourteen important and nonredundant features were found to be predictors of PC recurrence at a mean ± SD of 3.2 ± 2.2 years post-RP. Five-fold, ten-run cross-validation of the model containing these fourteen features yielded an area under the curve (AUC) of 0.89 ± 0.04 in the training set (n = 225). In comparison, the University of California San Fransisco Cancer of the Prostate Risk Assessment score (UCSF-CAPRA) and Memorial Sloan Kettering Cancer Center (MSKCC) Pre-Radical prostatectomy nomograms yielded AUC of 0.66 ± 0.05 and 0.67 ± 0.05, respectively (p < 0.01). When the radiomic model was applied to the test set (n = 26), AUC was 0.78; sensitivity, specificity, positive predictive value, and negative predictive value were 60%, 86%, 52%, and 89%, respectively. Accuracy in predicting PC recurrence was 81%. [ABSTRACT FROM AUTHOR]

Background: In light of multinational efforts to reduce helminthiasis, we evaluated whether there exist high-risk subpopulations for helminth infection. Such individuals are not only at risk of morbidity, but may be important parasite reservoirs and appropriate targets for disease control interventions. Methods/Principal Findings: We followed two longitudinal cohorts in Sichuan, China to determine whether there exist persistent human reservoirs for the water-borne helminth, Schistosoma japonicum, in areas where treatment is ongoing. Participants were tested for S. japonicum infection at enrollment and two follow-up points. All infections were promptly treated with praziquantel. We estimated the ratio of the observed to expected proportion of the population with two consecutive infections at follow-up. The expected proportion was estimated using a prevalence-based model and, as highly exposed individuals may be most likely to be repeatedly infected, a second model that accounted for exposure using a data adaptive, machine learning algorithm. Using the prevalence-based model, there were 1.5 and 5.8 times more individuals with two consecutive infections than expected in cohorts 1 and 2, respectively (p<0.001 in both cohorts). When we accounted for exposure, the ratio was 1.3 (p = 0.013) and 2.1 (p<0.001) in cohorts 1 and 2, respectively. Conclusions/Significance: We found clustering of infections within a limited number of hosts that was not fully explained by host exposure. This suggests some hosts may be particularly susceptible to S. japonicum infection, or that uncured infections persist despite treatment. We propose an explanatory model that suggests that as cercarial exposure declines, so too does the size of the vulnerable subpopulation. In low-prevalence settings, interventions targeting individuals with a history of S. japonicum infection may efficiently advance disease control efforts. Author Summary: Approximately 1 billion people are infected with one or more helminthes – a class of parasites that can impair physical, mental and economic development. We are interested in whether there exist groups who are repeatedly infected with helminthes over time in areas where treatment is ongoing. Such individuals may be at risk of morbidity and may also serve as parasite reservoirs, making them appropriate targets for disease control programs. We followed two cohorts in rural Sichuan, China in order to evaluate whether the same individuals were repeatedly infected with the water-borne helminth, Schistosoma japonicum. Each participant was tested for infection at enrollment and two follow-up points – all infections were promptly treated. We conducted detailed interviews to assess exposures to S. japonicum. We found infections repeatedly occurred in a subgroup of individuals and this clustering of infections was only partly explained by differences in exposure. This suggests some individuals may be particularly susceptible to S. japonicum infection. Further exploration of the interplay of exposure and susceptibility suggest that as exposure declines, so too does the fraction of the population vulnerable to infection. Helminth control programs that target people with a history of infection may efficiently reduce helminth infections and morbidity. [ABSTRACT FROM AUTHOR]

Background: In light of multinational efforts to reduce helminthiasis, we evaluated whether there exist high-risk subpopulations for helminth infection. Such individuals are not only at risk of morbidity, but may be important parasite reservoirs and appropriate targets for disease control interventions. Methods/Principal Findings: We followed two longitudinal cohorts in Sichuan, China to determine whether there exist persistent human reservoirs for the water-borne helminth, Schistosoma japonicum, in areas where treatment is ongoing. Participants were tested for S. japonicum infection at enrollment and two follow-up points. All infections were promptly treated with praziquantel. We estimated the ratio of the observed to expected proportion of the population with two consecutive infections at follow-up. The expected proportion was estimated using a prevalence-based model and, as highly exposed individuals may be most likely to be repeatedly infected, a second model that accounted for exposure using a data adaptive, machine learning algorithm. Using the prevalence-based model, there were 1.5 and 5.8 times more individuals with two consecutive infections than expected in cohorts 1 and 2, respectively (p,0.001 in both cohorts). When we accounted for exposure, the ratio was 1.3 (p = 0.013) and 2.1 (p,0.001) in cohorts 1 and 2, respectively. Conclusions/Significance: We found clustering of infections within a limited number of hosts that was not fully explained by host exposure. This suggests some hosts may be particularly susceptible to S. japonicum infection, or that uncured infections persist despite treatment. We propose an explanatory model that suggests that as cercarial exposure declines, so too does the size of the vulnerable subpopulation. In low-prevalence settings, interventions targeting individuals with a history of S. japonicum infection may efficiently advance disease control efforts. [ABSTRACT FROM AUTHOR]

[Display omitted] • A low-cost fabric strain sensor with a spiral wound structure has been developed. • The sensor exhibits excellent linear sensing characteristics. • The relative resistance changes are consistent under different loading rates. • The sensor possesses broad sensing range, fast response, and high stability. This work presents a strain sensor which is based on high ductility liquid metal coated on traditional commercial lastex yarns with a unique spiral wound structure. The prepared sensor exhibits exceptional flexibility, elasticity, softness of 0.23 N at 100% tensile strain. It is lightweight, weighing only 1 g/m, and boasts a low cost of less than 1 dollar. The sensor has the ability to detect both static and dynamic in-plane tension and normal pressure. It showcases remarkable sensing characteristics of good sensitivity (0.82 for tension, 12.6 N−1 for pressure), extraordinary linearity within 10–100% strain range with goodness-of-fit value over 0.997, rapid response of less than 10 ms, high resolution spanning up to 0.3%, and high durability over 100,000 repetitive cycles. A smart glove based on the proposed strain sensors achieves gesture recognitions using machine learning technique, with an accuracy of 100%. The outstanding performance of this sensor highlights its potential as an affordable and advanced candidate for wearable electronics, which can have wide-ranging applications in sports, healthcare, entertainment, education, and industry. [ABSTRACT FROM AUTHOR]

Locating diseases precisely from medical images, like ultrasonic and CT images, have been one of the most challenging problems in medical image analysis. In recent years, the vigorous development of deep learning models have greatly improved the accuracy in disease location on medical images. However, there are few artificial intelligent methods for identifying cholelithiasis and classifying gallstones on CT images, since no open source CT images dataset of cholelithiasis and gallstones is available for training the models and verifying their performance. In this paper, we build up the first medical image dataset of cholelithiasis by collecting 223846 CT images with gallstone of 1369 patients. With these CT images, a neural network is trained to “pick up” CT images of high quality as training set, and then a novel Yolo neural network, named Yolov3-arch neural network, is proposed to identify cholelithiasis and classify gallstones on CT images. Identification and classification accuracies are obtained by 10-fold cross-validations. It is obtained that our Yolov3-arch model is with average accuracy 92.7% in identifying granular gallstones and average accuracy 80.3% in identifying muddy gallstones. This achieves 3.5% and 8% improvements in identifying granular and muddy gallstones to general Yolo v3 model, respectively. Also, the average cholelithiasis identifying accuracy is improved to 86.50% from 80.75%. Meanwhile, our method can reduce the misdiagnosis rate of negative samples by the object detection model. [ABSTRACT FROM AUTHOR]

Background: A challenge preventing routine clinical implementation of Monte Carlo (MC)-based lung SBRT is the difficulty of reinterpreting historical outcome data calculated with inaccurate dose algorithms, because the target dose was found to decrease to varying degrees when recalculated with MC. The large variability was previously found to be affected by factors such as tumour size, location, and lung density, usually through sub-group comparisons. We hereby conducted a pilot study to systematically and quantitatively analyze these patient factors and explore accurate target dose conversion models, so that large-scale historical outcome data can be correlated with more accurate MC dose without recalculation.Methods: Twenty-one patients that underwent SBRT for early-stage lung cancer were replanned with 6MV 360° dynamic conformal arcs using pencil-beam (PB) and recalculated with MC. The percent D95 difference (PB-MC) was calculated for the PTV and GTV. Using single linear regression, this difference was correlated with the following quantitative patient indices: maximum tumour diameter (MaxD); PTV and GTV volumes; minimum distance from tumour to soft tissue (dmin); and mean density and standard deviation of the PTV, GTV, PTV margin, lung, and 2 mm, 15 mm, 50 mm shells outside the PTV. Multiple linear regression and artificial neural network (ANN) were employed to model multiple factors and improve dose conversion accuracy.Results: Single linear regression with PTV D95 deficiency identified the strongest correlation on mean-density (location) indices, weaker on lung density, and the weakest on size indices, with the following R(2) values in decreasing orders: shell2mm (0.71), PTV (0.68), PTV margin (0.65), shell15mm (0.62), shell50mm (0.49), lung (0.40), dmin (0.22), GTV (0.19), MaxD (0.17), PTV volume (0.15), and GTV volume (0.08). A multiple linear regression model yielded the significance factor of 3.0E-7 using two independent features: mean density of shell2mm (P = 1.6E-7) and PTV volume (P = 0.006). A 4-feature ANN model slightly improved the modeling accuracy.Conclusion: Quantifiable density features were proposed, replacing simple central/peripheral location designation, which showed strong correlations with PB-to-MC target dose conversion magnitude, followed by lung density and target size. Density in the immediate outer and inner areas of the PTV showed the strongest correlations. A multiple linear regression model with one such feature and PTV volume established a high significance factor, improving dose conversion accuracy. [ABSTRACT FROM AUTHOR]

Class evolution, the phenomenon of class emergence and disappearance, is an important research topic for data stream mining. All previous studies implicitly regard class evolution as a transient change, which is not true for many real-world problems. This paper concerns the scenario where classes emerge or disappear gradually. A class-based ensemble approach, namely Class-Based ensemble for Class Evolution (CBCE), is proposed. By maintaining a base learner for each class and dynamically updating the base learners with new data, CBCE can rapidly adjust to class evolution. A novel under-sampling method for the base learners is also proposed to handle the dynamic class-imbalance problem caused by the gradual evolution of classes. Empirical studies demonstrate the effectiveness of CBCE in various class evolution scenarios in comparison to existing class evolution adaptation methods. [ABSTRACT FROM PUBLISHER]