Search

As a foundation for supporting downstream tasks, the named entity recognition (NER) task in the field of traditional Chinese medicine (TCM) is gaining increasing attention. However, the TCM domain suffers from a lack of labeled data and a high annotation error rate. To solve this problem, we propose an abductive learning framework for named TCM entity recognition called ABL-TCM. First, we design a series of methods to mitigate the negative impact of mislabeling on the effectiveness of the constructed model and to enhance the focus of the training process on challenging entities. Second, we offer a flexible approach for incorporating external unsupervised data. Finally, we propose a label correction mechanism based on abductive learning to ensure that only reliable data are included in the training process. The results of several experiments prove the effectiveness of ABL-TCM, which achieves the best results on a TCM-NER dataset.

In the voltage source converter based high-voltage direct current (VSC-HVDC) grids, fast and reliable protections are the key technologies. The traditional protection schemes are easily affected by fault resistance, line distributed capacitance, etc. Meanwhile, the influence of fault current limiting strategy (FCLS) has not been fully considered. In this paper, the fault characteristics under FCLS and the feasibility of traditional travelling wave protections are analyzed. To improve the reliability and sensibility, a similarity comparison based pilot protection scheme is proposed, which focuses on the relationship between the fault characteristics and the state of the protected transmission line, with the establishment of a precise frequency-dependent transmission line model. The criteria based on the similarity comparison calculated by cross-wavelet can identify the fault effectively. Meanwhile, the protection scheme can also endure the influence of error synchronization. Finally, the protection performance is verified in the PSCAD/EMTDC under different fault conditions.

In the past few years, sparse subspace clustering (SSC) has gained many studies and found wide applications. However, SSC suffers from the limitation in scalability. Furthermore, current SSC methods could hardly tackle stream data where the structure of subspaces may change along time. In this paper, we propose a novel method to extend SSC to stream data (StreamSSC). Our method is based on maintaining a small subset of representatives to characterize the structure of the underlying subspaces during stream data. StreamSSC is efficient in both computation and memory. Experimental results on both synthetic and real-world streams demonstrate the effectiveness of StreamSSC. For efficiency, StreamSSC is faster than existing online subspace clustering methods by roughly a magnitude.

As an effective technique to learn low-dimensional node features in complicated network environment, network embedding has become a promising research direction in the field of network analysis. Due to the virtues of better interpretability and flexibility, matrix factorization based methods for network embedding have received increasing attentions. However, most of them are inadequate to learn more complicated hierarchical features hidden in complex networks because of their mechanisms of single-layer factorization structure. Besides, their original feature matrices used for factorization and their robustness against noises also need to be further improved. To solve these problems, we propose a novel network embedding method named DRNMF (deep robust nonnegative matrix factorization), which is formed by multi-layer NMF learning structure. Meanwhile, DRNMF employs the combination of high-order proximity matrices of the network as the original feature matrix for the factorization. To improve the robustness against noises, we use ℓ2,1 norm to devise the objective function for the DRNMF network embedding model. Effective iterative update rules are derived to resolve the model, and the convergence of these rules are strictly proved. Moreover, we introduce a pre-training strategy to improve the efficiency of convergence. Extensive experiments on several benchmarks of complex networks demonstrate that our proposed method DRNMF is effective and has better performance than the state-of-the-art matrix factorization based methods for network embedding.

Adequate evidence has shown that miRNA-disease interactions are strongly involved in the pathological processes of complex human diseases. However, it is commonly time-consuming and labor-intensive to utilize laboratory biological experiments to reveal unknown miRNA-disease pairs. Since the previously proposed calculation model has more or fewer deficiencies, we developed the semi-supervised method called Hessian Regularized Non-negative Matrix Factorization Method for miRNA-disease Association prediction (HRNMFMDA). This model introduced Hessian regularization into the NMF framework to preserve the local manifold information and increased an $l_{2,1}$ -norm penalty term to ensure the feature selection of the coding matrix and an approximate orthogonal constraint to obtain discriminative information. In the model performance evaluation, HRNMFMDA outperformed eight existing models, achieving AUC values of 0.9074, 0.8618, and 0.9044+/-0.0080 in the global leave-one-out cross-validation (LOOCV), local LOOCV and 5-fold cross-validation, respectively. In addition, we applied HRNMFMDA to several high-incidence human carcinomas via three kinds of case studies. Almost all predicted miRNAs were confirmed by external databases derived from experimental literature. Therefore, the conclusion can be drawn that HRNMFMDA is reliable for revealing uncovered miRNA-disease pairs.

Resource scheduling problems (RSPs) in cloud manufacturing (CMfg) often manifest as dynamic scheduling problems in which scheduling strategies depend on real-time environments and demands. Generally, multiple resources in the CMfg scheduling process cause difficulties in system modeling. To solve this problem, we propose Sharer, a deep reinforcement learning (DRL)-based method that converts scheduling problems with multiple resources into one learning target and learns effective strategies automatically. Our preliminary results show that Sharer is comparable to the latest heuristics, adapts to different conditions, converges quickly, and subsequently learns wise strategies.

Perovskite quantum dots (QDs) attracted a lot of attention recently due to their high light conversion efficiency, and high color purity. However, perovskite materials are sensitive to heat. High temperature will cause serious quenching problem of the perovskite QDs. In order to solve the QDs thermal quenching problem without destroying their optical characters. A circulating quantum dots colloidal solution converter (CQD-converter) was developed. A closed loop pump tube coated with Teflon was utilized to prevent perovskite QDs from the external environment. A peristaltic pump was utilized to drive the QDs colloidal solution through a Peltier cooler for dissipating the heat of the QDs colloidal solution. The lifetime of the perovskite QDs can be significantly improved under a 450nm laser excitation. The effects of temperature of the Peltier cooler and flow velocity on the optical properties of QDs were also investigated. Compared to the static quantum dots colloidal solution (SQD-converter), which quenched after 7 min, the optical properties were maintained after 600 min for the CQD-converter.

The inorganic halide perovskite quantum dots (QDs) have been considered as a promising substitute for white light-emitting diodes (WLEDs). In this paper, the green CsPbBr3 QDs and red K2SiF6:Mn4+ (KSF) phosphor were used to fabricate the conversion layers. Because the location of the layers is fundamental to the absorption priority of blue light, the location of KSF and QDs were controlled and the QDs-up type and QDs-down type WLEDs were made. The optical power, luminous efficiency, CCE, and luminous intensity in the middle of QDs-up type are 13.83 mW, 13.54 lm/W, 11.94%, and 2.65 cd, meaning 24.26%, 25.72%, 2.63%, and 23.83% higher than those of QDs-down type, respectively. In addition, the QDs-up type has a lower correlated color temperature (CCT) shift of 734 K and a decreased highest temperature of 56.8° (51.5% and 14.9% lower). These key property differences indicate that the QDs-up type is more suitable for the application in display and backlight. In order to explore the reasons for these differences, the emission spectra, CCE, reflection rate, absorption rate, and temperature curves of QDs or KSF films were also analyzed, which provided a better understanding of designing package structures.

Mutation-based greybox fuzzing is a highly effective and widely used technique to find bugs in software. Provided initial seeds, fuzzers continuously generate test cases to test the software by mutating a seed input. However, the majority of them are “invalid” because the mutation may destroy the format of the seeds. In this paper, we present a knowledge-learn evolutionary fuzzer based on AFL, which is called LearnAFL. LearnAFL does not require any prior knowledge of the application or input format. Based on our format generation theory, LearnAFL can learn partial format knowledge of some paths by analyzing the test cases that exercise the paths. Then LearnAFL uses these format information to mutate the seeds, which is efficient to explore deeper paths and reduce the test cases exercising high-frequency paths than AFL. We compared LearnAFL with AFL and some other state-of-the-art fuzzers on ten real-world programs. The result showed that LearnAFL could reach branch coverage 120% and 110% of that of AFL and FairFuzz, respectively. LearnAFL also found 8 unknown vulnerabilities in GNU Binutils, Libpng and Gif2png, all of which have been reported to the vendors. Besides, we compared the format information learned from the initial seed of an ELF file with a format standard of ELF files. The result showed that LearnAFL learns about 64% part of the file format without any prior knowledge.

Network embedding, aiming to learn low-dimensional representations of nodes in networks, is very useful for many vector-based machine learning algorithms and has become a hot research topic in network analysis. Although many methods for network embedding have been proposed before, most of them are unsupervised, which ignores the role of prior information available in the network. In this paper, we propose a novel method for network embedding using semi-supervised kernel nonnegative matrix factorization (SSKNMF), which can incorporate prior information and thus to learn more useful features from the network through introducing kernel methodology. Besides, it can improve robustness against noises by using the objective function based on L2,1 norm. Efficient iterative update rules are derived to resolve the network embedding model using the SSKNMF, and the convergence of these rules are strictly proved from the perspective of mathematics. The results from extensive experiments on several real-world networks show that our proposed algorithm is effective and has better performance than the existing representative methods.

Academic procrastination has been reported affecting students' performance in computer-supported learning environments. Studies have shown that students who demonstrate higher procrastination tendencies achieve less than the students with lower procrastination tendencies. It is important for a teacher to be aware of the students' behaviors especially their procrastination trends. EDM techniques can be used to analyze data collected through computer-supported learning environments and to predict students' behaviors. In this paper, we present an algorithm called students' academic performance enhancement through homework late/non-submission detection (SAPE) for predicting students' academic performance. This algorithm is designed to predict students with learning difficulties through their homework submission behaviors. First, students are labeled as procrastinators or non-procrastinators using k-means clustering algorithm. Then, different classification methods are used to classify students using homework submission feature vectors. We use ten classification methods, i.e., ZeroR, OneR, ID3, J48, random forest, decision stump, JRip, PART, NBTree, and Prism. A detailed analysis is presented regarding performance of different classification methods for different number of classes. The analysis reveals that in general the prediction accuracy of all methods decreases with increase in the number of classes. However, different methods perform best or worst for different number of classes.

Micro-scale patterned arrays and nano-scale rough morphology are promising for improving the light-extraction performance of GaN-based thin film light-emitting diodes (TFLEDs), while the light-extraction mechanisms of the multiscale architectures combining these two structures have not been investigated yet. In this report, we have adopted a pattern transfer and wet etching combined method to fabricate multiscale patterned arrays with rough morphology (msPARM) on n-GaN layers for TFLEDs and investigated their light-extraction mechanisms by the finite-difference time domain and ray-tracing combined method. The results show that the TFLEDs achieve the maximum radiant efficacy using the msPARM with an etching time of 8 min, which is increased by 16.3% and 1.7% compared with that achieved using only the patterned arrays or only the rough morphology, respectively. Most importantly, optical simulation reveals that the msPARM can provide a high transmittance for light with large emission angles from the active region using the inclined surface of micro-scale concave cones, while effectively suppressing the reflection loss for light with small emission angles using the scattering effect of nano-scale rough morphology, resulting in enhancing the light-extraction of the TFLEDs. Consequently, this study can provide a better understanding to design the multiscale structures for achieving high efficiency LEDs.

A novel flexible woven heat sink (FWHS) composed of sintered copper braided belts was proposed here to meet the heat dissipation requirements of the flexible radiator installed in a narrow space. The sintering treatment was employed to ameliorate the tensile and bending properties as well as the thermal conductivity of the FWHS. The thermal resistance network model and the temperature field analysis in cases of natural convection and forced convection were presented to evaluate the thermal performance of the FWHS. The results show that the FWHS sintered at 850 °C with the expanded shape at a compressed height of 55 mm can reduce the temperature of heat source by 17.1% compared to the uncompressed FWHS, which demonstrates its outstanding advantage of maintaining high efficiency of heat dissipation under compressive deformation.

Mobile cloud computing provides applications of cloud computing technology on mobile Internet. In the era of the 5G networks, mobile cloud computing offers the possibility of heavy computational in parallel through different mobile terminals. This also offers us to use mobile cloud computing to analyze fluctuations in user buying incentives over the course of a year. In various business scenarios, the purchase incentive of the public is often considered as being controlled by multiple factors and changes dramatically during holiday seasons. This paper proposed an evolutionary algorithm for monitoring purchase incentive of the public which can be distributed in mobile cloud computing. In this paper, we demonstrate that the public's shopping behavior of a commodity is a consequence of a collective behavior propagating through a social network and can be modeled as a simple ODEs model. Then, based on this new model, we develop an extended differential evolution algorithm that is deployed on mobile terminals and combined with the 5G network computing to estimate public incentives from historical sales datasets. The results suggest that this method can successfully monitor the public's purchase incentive. In addition, the overall purchase incentive changes dramatically during holiday seasons but fluctuates with similar patterns every year.

Recently, deep learning has gained great popularity in the area of recommender systems. Various combinations of deep learning, collaborative recommendation and content-based recommendation have occurred. However, as one of the three most significant recommendation techniques, hybrid recommendation has little cooperation with deep learning. Besides, most current deep hybrid models only incorporate two simple recommendation methods together in post-fusion, leaving massive space for further exploration of better combinations. In this paper, we apply deep learning to hybrid recommendation, proposing a deep hybrid recommendation model DMFL (Deep Metric Factorization Learning). In DMFL, we combine deep learning with improved machine learning models to learn the interaction between users and items from multiple perspectives. Such deep hybrid learning helps to reflect the user preference more comprehensively and strengthen model's ability of generalization. We also propose a more accurate method of user feature representation, taking both long-term static characteristics and short-term dynamic interest changes of users into consideration. Furthermore, thorough experiments have been conducted on real-world datasets, which strongly proves the effectiveness and efficiency of the proposed model.

In recent years, network embedding has attracted more and more attention due to its effectiveness and convenience to compress the network structured data. In this paper, we propose a community-based variational autoencoder (ComVAE) model to learn network embedding, which consists of a community detection module and a deep learning module. In the proposed model, both community information and deep learning techniques are utilized to learn low-dimensional vertex representations. First, community information reveals an implicit relationship between vertices from a global view, which can be a supplement to local information and help to improve the embedding quality. To obtain the community information, community detection algorithms are utilized as a module and the modularization design makes the model more flexible. Second, deep learning techniques can not only integrate and preserve the information from both local and global views efficiently but also strengthen the robustness of vertex representations. To demonstrate the performance of our model, extensive experiments are conducted in four downstream tasks, namely, network reconstruction, node classification, link prediction, and visualization. The experimental results show that our model outperforms the state-of-the-art approaches to real-world datasets.

To alleviate the problem caused by the sparsity of network structure which is often the case in large-scale network, attributed network embedding has attracted an increasing amount of attention. Some existing attributed network embedding models integrate linkage structure and node attribute by adding a consistency criterion on the structure representation and attribute representation, whereby the similarity between them can be ensured. However, to enforce the structure and attribute representations to be similar may cause the information distortion due to the inherent difference between these two kinds of information. Additionally, the existing models mainly focus on learning the inter-relation between structure and attribute information, while the intra-characteristic of each information is ignored, leading to the information loss. To address the above two problems, we propose a novel model named Inter-Intra Information Preserving Attributed Network Embedding (IINE) to effectively learn the node representations of attributed network, which can not only capture the inter-relation between structure and attribute information with less information distortion but also effectively preserve the intra-characteristic of each information. The proposed model is composed of a primary model named coupled autoencoder and two auxiliary models named structure miner and attribute miner. The coupled autoencoder trains the node representation by smoothly combining both structure and attribute information, while the structure miner and attribute miner are utilized to further mine the intra-characteristic from the corresponding information so as to assist the primary model. The Extensive experiments are conducted on seven real-world datasets, and the results confirm the superior performance of IINE over several state-of-the-art models.

Phosphor-converted light-emitting diode (pcLED) is one of the most promising methods to generate white light. The interaction of the excitation light with the phosphor in pcLED is a complicated phenomenon, combining light absorption, light conversion, and scattering. An effective and efficient way to predict the optical behavior of pcLED is using the bidirectional scattering distribution function (BSDF). In this paper, the BSDF for pcLEDs with arbitrary phosphor concentration and thickness was established. The measured integrating energy and angular distributions of BSDF were successfully fitted by the interpolation method and the ABg model, respectively, realizing the calculation of BSDF values for arbitrary phosphor configurations. The proposed model was applied to a remote pcLED using an arbitrary phosphor configuration in Monte Carlo simulation. The output radiant power and the angular distributions of the blue and yellow light show good agreements between experiments and simulations; the errors of radiant power were all less than 4%, validating the effectiveness of the model. Consequently, this study provides a simple and useful tool to precisely predict the optical performance of pcLEDs with arbitrary concentration and thickness, it is essential for white LED design and fabrication.