查看更多>>摘要：Genetic Programming (GP) has been successfully applied to image classification and achieved promising results. However, most existing methods either address binary image classification tasks only or need a predefined classifier to perform multi-class image classification while using GP for feature extraction. This limits their flexibility since it is unknown which combinations of classifiers and features are the most effective for an image classification task. Furthermore, high image variations increase the difficulty of feature extraction and image classification. This paper proposes a GP approach with a new program representation, new functions, and new terminals. The new approach can conduct feature extraction, feature construction, and classification, automatically and simultaneously. It can extract and construct informative image features, select a suitable classification algorithm instead of relying on a predefined classifier, and perform classification for binary and multi-class image classification tasks. In addition, this paper develops a new mutation operator based on fitness of population for dynamically adjusting the size of the evolved GP programs. The experimental results on eight datasets with different variations and difficulties show that the proposed approach achieves higher classification accuracy than most of the benchmark methods. Further analysis shows that the GP evolved programs have appropriate tree sizes and potentially high interpretability.
查看更多>>摘要：This paper focuses on the problem of optimizing the kernel and penalty parameters for SVM classifiers with Gaussian kernel. To reduce the computational overhead of inter-cluster distance in the feature space (ICDF) with a large number of candidate discretized values in a large interval in previous researches, in this paper, the new inter-cluster induced distance in the feature space (ICIDF) is proposed to guide the kernel parameter selection of SVMs, and the theorem that the ICIDF is a positive strictly unimodal function about Gaussian kernel parameter is firstly presented. Then, a fast parameter optimization approach including two stages is presented for SVMs according to this theorem. In the first stage, a modified golden section algorithm (MGSA) is proposed to obtain a shrunk value interval for kernel parameter in small amount of ICIDF calculations. In the second stage, a differential evolutionary algorithm (BBDE or SADE) is applied to select the best parameter combination for SVM in the shrunk interval of kernel parameter obtained by MGSA and a given interval of penalty parameter. Experiments for benchmark datasets illustrate that the training time of SVM models can significantly shortened by our approach, while the testing accuracy of the trained SVMs is competitive.
查看更多>>摘要：Opinion target extraction (OTE) is the extraction of explicit expressions related to entity aspects interpreted with subjective attributive words in the review sentences using supervised or rule-based approaches. Despite the constraints of syntactic-based relation rules, rule-based approaches can be domain-independently implemented. Although supervised approaches yield better results, more costly due to requiring a large number of labeled samples. This study proposes an unsupervised (rule-based) OTE approach with novel methods and extended rule-based techniques to overcome the aforementioned issues. In this study, first, a novel pattern-based text pre-processing method is proposed to eliminate punctuations that are incompatible with determinative group rules patterns. Then, implemented syntactic-based relation rules on the dependency relation graph are extended with new auxiliary features to extract multi-word expressions which modify each other. The majority voting method is used for optimizing the performance of outputs. Finally, the effectiveness of the proposed approach was tested on a restaurant review dataset. The experimental results show that the proposed approach outperforms all unsupervised approaches. Additionally, it gives comparable results with the supervised approaches, revealing the effectiveness of the proposed approach.
Del Ser J.Casillas-Perez D.Cornejo-Bueno L.Salcedo-Sanz S....
24页
查看更多>>摘要：In the last few years, methods falling within the family of randomization-based machine learning models have grasped a great interest in the Artificial Intelligence community, mainly due to their excellent balance between performance in prediction problems and their computational efficiency. The use of these models for prediction problems related to renewable energy sources has been particularly notable in recent times, including different ways in which randomization is considered, their hybridization with other modeling techniques and/or their multi-layered (deep) and ensemble arrangement. This manuscript comprehensively reviews the most important features of randomization-based machine learning methods, and critically examines recent evidences of their application to renewable energy prediction problems, focusing on those related to solar, wind, marine/ocean and hydro-power renewable sources. Our study of the literature is complemented by an extensive experimental setup encompassing three real-world problems dealing with solar radiation prediction, wind speed prediction in wind farms and hydro-power energy. In all these problems randomization-based methods are reported to achieve a better predictive performance than their corresponding state-of-the-art solutions, while demanding a dramatically lower computational effort for its learning phases. Finally, we pause and reflect on important challenges faced by these methods when applied to renewable energy prediction problems, such as their intrinsic epistemic uncertainty, or the need for explainability. We also point out several research opportunities that arise from this vibrant research area.
查看更多>>摘要：In recent years, with the rapid development of convolutional neural networks (CNNs), a series of computer vision tasks have been solved. However, scene recognition is still a difficult and challenging problem due to the complexity of scene images. With the emergence of large-scale scene datasets, a single representation generated by a plain CNN is no longer discriminative enough to describe massive scene images. Therefore, in this paper, we propose a comprehensive representation for scene recognition, including enhanced global scene representation, local salient scene representation, and local contextual object representation. We use two pretrained CNNs to extract original feature maps to construct the multiple representations. Specifically, we adopt class activation mapping (CAM) to find salient regions and extract local scene features and employ a bidirectional long short-term module (LSTM) to encode contextual information of objects existing in a scene. In addition, the multiple representations are generated by an end-to-end trainable model, which we call MRNet (multiple representation network). Experimental results on three publicly available scene recognition datasets demonstrate that our proposed model is superior to state-of-the-art models.
查看更多>>摘要：Neuro-fuzzy models have been used to predict runoff from rainfall, a hydrological phenomenon associated with a degree of uncertainty. However, rainfall can be measured from different meteorological stations, and runoff forecasting can be biased. Thus, the aim of this work is to propose a new stacking neuro-fuzzy framework for predicting runoff from physically distributed meteorological stations. As a method to estimate single one-day-ahead runoff and as a stacking approach, the Self-Identification Neuro-fuzzy Inference model (SINFIM) and Self-Organizing Neuro-fuzzy Inference System (SONFIS) were developed, respectively. As a case study, data from two Chilean watersheds (the Diguillín River (?uble region) and Colorado River (Maule region)) and average daily runoff and average daily rainfall recorded over eighteen years were collected from the Chilean Directorate of Water Resources (DGA). The experimental results show good adjustment in the single forecasting of runoff with meteorological stations showing adjustment and efficiency indexes of greater than 80% in the validation set and being able to efficiently predict both high and low runoff values. However, better results were obtained with the stacking model with values being higher than single runoff predictions and those of state-of-art approaches. Therefore, the general framework proposed represents a good approach for forecasting runoff since it can improve predictions and generate more accurate runoff values than single models.
查看更多>>摘要：Medical image fusion of images obtained via different modes can expand the inherent information of original images, whereby the fused image has a superior ability to display details than the original sub-images, to facilitate diagnosis and treatment selection. In medical image fusion, an inherent challenge is to effectively combine the most useful information and image details without information loss. Despite the many methods that have been proposed, the effective retention and presentation of information proves challenging. Therefore, we proposed and evaluated a novel image fusion method based on the discrete stationary wavelet transform (DSWT) and radial basis function neural network (RBFNN). First, we analyze the details or feature information of two images to be processed by DSWT by using two-level decomposition to separate each image into seven parts, comprising both high-frequency and low-frequency sub-bands. Considering the gradient and energy attributes of the target, we substituted the pending parts in the same position in the two images by using the proposed enhanced RBFNN. The input, hidden, and output layers of the neural network comprised 8, 40, and 1 neuron(s), respectively. From the seven neural networks, we obtained seven fused parts. Finally, through inverse wavelet transform, we obtained the final fused image. For the neural network training method, the hybrid adaptive gradient descent algorithm (AGDA) and gravitational search algorithm (GSA) were implemented. The final experimental results revealed that the novel method has significantly better performance than the current state-of-the-art methods.
查看更多>>摘要：Traffic congestion is a vital factor hindering travel. As such, developing a reliable traffic congestion early warning system is essential for providing traffic condition supervision and programming. However, previous research has rarely focused on traffic flow characteristics or on providing comprehensive assessments, resulting in poor warning performances. In this study, an innovative traffic congestion early warning system is proposed, comprising point forecasting, characteristic estimate, interval prediction, and comprehensive assessment. In the characteristic assessment phase, eight common statistical distributions are used to fit the characteristics of an original traffic flow parameter series in a training set, and the best fitting results are considered as the basis for building a prediction interval. An extreme learning machine combined with a modified multi-objective dragonfly optimization algorithm and variational mode decomposition is constructed in the point forecasting phase to provide accurate and stable traffic flow parameter forecasting results; two different strategies are used to establish the prediction interval, so as to conduct interval forecasting based on different types of uncertainty information (probability distribution information or known interval information). Extenics evaluation theory is then used in the comprehensive assessment phase to evaluate the traffic congestion level. Simulations of traffic flow parameter series, including simulations of the road density, road occupancy, and average velocity, reveal that the proposed early warning system demonstrates powerful abilities based on its precision and stability. The mean absolute percentage error (MAPE) values of the traffic flow parameters for the three datasets are 3.6265%, 3.7203%, and 4.5100%, respectively. The forecasting accuracy for the traffic congestion level is more than 97% for both point and interval prediction. Thus, this approach can be widely used for personal traffic route planning and the unified management of governmental traffic conditions.
查看更多>>摘要：Temporal and spatial information about scholars can be found in their academic papers. Examining the footprints of scholars’ careers can help us understand the course of their growth, potential collaborations for future research, and trends in the flow of talent. Although a great deal of research has been conducted in related fields, the challenge of accurately constructing scholars’ career trajectories from redundant and noisy academic data is far from resolved. To address this problem, a unified framework called ATrajRN that employs AMiner academic data is proposed for the first time. To accurately obtain scholars’ geographic location information from their research achievements, this study introduces an algorithm called Positioning based on Academic Achievements of Scholars (PAAS), which aims to make the most of academic data and the characteristics of different maps. To avoid the interference of data redundancy, this paper proposes a statistical feature-based method to find the most reliable career trajectories by some state-of-the-art approaches. To restore the continuously scholars’ career trajectories, this paper offers the trajectory generation algorithm based on the output from the previous step. Experiments and systematic analysis shows that the proposed novel method could achieve approximately 80% accuracy – an increase of approximately 10% – manifestly outperform the baseline method. Lastly, based on this work, we develop a system for understanding scholars’ trajectories through analysis and visualization, and we investigate the migration characteristics of typical scholar groups.
查看更多>>摘要：The minimum vertex cover problem belongs to a NP- complete problem, which is difficult to obtain the near-optimal solution in the polynomial time range using classical algorithms. In this paper, a quantum circuit solution scheme based on the quantum approximate optimization algorithm is presented for the minimum vertex cover problem. Firstly, the quantum Ising model and Hamiltonian of the problem are obtained based on the Ising model corresponding to the problem, which is quantized by the rotation operator and Pauli operator. Secondly, the parametric unitary transformation with the initial Hamiltonian and the problem Hamiltonian as the generator is obtained respectively. Through the alternating evolution of two parametric unitary transformations, the final quantum state and the problem Hamiltonian expectation are derived. In the process of evolution, the parameters in the parametric unitary transformations which are optimized by the classical processor can adjust the problem Hamiltonian expectation, so as to improve the probability of the problem solution. Then, the initial state of the algorithm and the quantum logic gate corresponding to the parametric unitary transformation are derived to generate the quantum circuit which can be implemented on the quantum computer. Simulation results show that the scheme can obtain the problem solution with high probability in polynomial time, realizes exponential acceleration, and has certain feasibility, effectiveness and innovation.
NSTL
Elsevier