Knowledge discovery method for feature-decision level fusion of multiple classifiers

To improve the performance of the multiple classifier system, a new method of feature-decision level fusion is proposed based on knowledge discovery. In the new method, the base classifiers operate on different feature spaces and their types depend on different measures of between-class separability. The uncertainty measures corresponding to each output of each base classifier are induced from the established decision tables （DTs） in the form of mass function in the Dempster-Shafer theory （DST）. Furthermore, an effective fusion framework is built at the feature-decision level on the basis of a generalized rough set model and the DST. The experiment for the classification of hyperspectral remote sensing images shows that the performance of the classification can be improved by the proposed method compared with that of plurality voting （PV）.

Emotional dialog generation via multiple classifiers based on a generative adversarial network

Background Human-machine dialog generation is an essential topic of research in the field of natural language processing.Generating high-quality,diverse,fluent,and emotional conversation is a challenging task.Based on continuing advancements in artificial intelligence and deep learning,new methods have come to the forefront in recent times.In particular,the end-to-end neural network model provides an extensible conversation generation framework that has the potential to enable machines to understand semantics and automatically generate responses.However,neural network models come with their own set of questions and challenges.The basic conversational model framework tends to produce universal,meaningless,and relatively"safe"answers.Methods Based on generative adversarial networks(GANs),a new emotional dialog generation framework called EMC-GAN is proposed in this study to address the task of emotional dialog generation.The proposed model comprises a generative and three discriminative models.The generator is based on the basic sequence-to-sequence(Seq2Seq)dialog generation model,and the aggregate discriminative model for the overall framework consists of a basic discriminative model,an emotion discriminative model,and a fluency discriminative model.The basic discriminative model distinguishes generated fake sentences from real sentences in the training corpus.The emotion discriminative model evaluates whether the emotion conveyed via the generated dialog agrees with a pre-specified emotion,and directs the generative model to generate dialogs that correspond to the category of the pre-specified emotion.Finally,the fluency discriminative model assigns a score to the fluency of the generated dialog and guides the generator to produce more fluent sentences.Results Based on the experimental results,this study confirms the superiority of the proposed model over similar existing models with respect to emotional accuracy,fluency,and consistency.Conclusions The proposed EMC-GAN model is capable of generating consistent,smooth,and fluent dialog that conveys pre-specified emotions,and exhibits better performance with respect to emotional accuracy,consistency,and fluency compared to its competitors.

Novel ensemble learning based on multiple section distribution in distributed environment

Because most ensemble learning algorithms use the centralized model, and the training instances must be centralized on a single station, it is difficult to centralize the training data on a station. A distributed ensemble learning algorithm is proposed which has two kinds of weight genes of instances that denote the global distribution and the local distribution. Instead of the repeated sampling method in the standard ensemble learning, non-balance sampling from each station is used to train the base classifier set of each station. The concept of the effective nearby region for local integration classifier is proposed, and is used for the dynamic integration method of multiple classifiers in distributed environment. The experiments show that the ensemble learning algorithm in distributed environment proposed could reduce the time of training the base classifiers effectively, and ensure the classify performance is as same as the centralized learning method.

Recognition of Characters by Adaptive Combination of Classifiers

In this paper, the visual feature space based on the long Horizontals, the long Verticals, and the radicals are given. An adaptive combination of classifiers, whose coefficients vary with the input pattern, is also proposed. Experiments show that the approach is promising for character recognition in video sequences.

Dynamic weighted voting for multiple classifier fusion:a generalized rough set method

To improve the performance of multiple classifier system, a knowledge discovery based dynamic weighted voting （KD-DWV） is proposed based on knowledge discovery. In the method, all base classifiers may be allowed to operate in different measurement/feature spaces to make the most of diverse classification information. The weights assigned to each output of a base classifier are estimated by the separability of training sample sets in relevant feature space. For this purpose, some decision tables （DTs） are established in terms of the diverse feature sets. And then the uncertainty measures of the separability are induced, in the form of mass functions in Dempster-Shafer theory （DST）, from each DTs based on generalized rough set model. From the mass functions, all the weights are calculated by a modified heuristic fusion function and assigned dynamically to each classifier varying with its output. The comparison experiment is performed on the hyperspectral remote sensing images. And the experimental results show that the performance of the classification can be improved by using the proposed method compared with the plurality voting （PV）.

Diversity-accuracy assessment of multiple classifier systems for the land cover classification of the Khumbu region in the Himalayas

Land cover classification of mountainous environments continues to be a challenging remote sensing problem,owing to landscape complexities exhibited by the region.This study explored a multiple classifier system(MCS)approach to the classification of mountain land cover for the Khumbu region in the Himalayas using Sentinel-2 images and a cloud-based model framework.The relationship between classification accuracy and MCS diversity was investigated,and the effects of different diversification and combination methods on MCS classification performance were comparatively assessed for this environment.We present ten MCS models that implement a homogeneous ensemble approach,using the high performing Random Forest(RF)algorithm as the selected classifier.The base classifiers of each MCS model were developed using different combinations of three diversity techniques:(1)distinct training sets,(2)Mean Decrease Accuracy feature selection,and(3)‘One-vs-All’problem reduction.The base classifier predictions of each RFMCS model were combined using:(1)majority vote,(2)weighted argmax,and(3)a meta RF classifier.All MCS models reported higher classification accuracies than the benchmark classifier(overall accuracy with 95% confidence interval:87.33%±0.97%),with the highest performing model reporting an overall accuracy(±95% confidence interval)of 90.95%±0.84%.Our key findings include:(1)MCS is effective in mountainous environments prone to noise from landscape complexities,(2)problem reduction is indicated as a stronger method over feature selection in improving the diversity of the MCS,(3)although the MCS diversity and accuracy have a positive correlation,our results suggest this is a weak relationship for mountainous classifications,and(4)the selected diversity methods improve the discriminability of MCS against vegetation and forest classes in mountainous land cover classifications and exhibit a cumulative effect on MCS diversity for this context.

Ensemble Classifier Design Based on Perturbation Binary Salp Swarm Algorithm for Classification

Multiple classifier system exhibits strong classification capacity compared with single classifiers,but they require significant computational resources.Selective ensemble system aims to attain equivalent or better classification accuracy with fewer classifiers.However,current methods fail to identify precise solutions for constructing an ensemble classifier.In this study,we propose an ensemble classifier design technique based on the perturbation binary salp swarm algorithm(ECDPB).Considering that extreme learning machines(ELMs)have rapid learning rates and good generalization ability,they can serve as the basic classifier for creating multiple candidates while using fewer computational resources.Meanwhile,we introduce a combined diversity measure by taking the complementarity and accuracy of ELMs into account;it is used to identify the ELMs that have good diversity and low error.In addition,we propose an ECDPB with powerful optimizing ability;it is employed to find the optimal subset of ELMs.The selected ELMs can then be used to forman ensemble classifier.Experiments on 10 benchmark datasets have been conducted,and the results demonstrate that the proposed ECDPB delivers superior classification capacity when compared with alternative methods.

Multivariety and multimanufacturer drug identification based on near-infrared spectroscopy and recurrent neural network

Near-infrared(NIR)spectral analysis,which has the advantages of rapidness,nondestruction and high-efficiency,is widely used in the detection of feed,food and mineral.In terms of qualitative identification,it can also be used for the discriminant analysis of medicines.Long short-term memory(LSTM)neural network,bidirectional long short-term memory(BiLSTM)neural network and gated recurrent unit(GRU)network are variants of the recurrent neural network(RNN).The potential relationship between nonlinear features learned from the sequence by these variants is used to complete the missions infields such as natural language processing,signal classification and video analysis.Since the effect of these variants in drug identification is still to be studied,this paper constructs a multiclassifier of these three variants,using compoundα-keto acid tablets produced by four manufacturers and repaglinide tablets produced by five manufacturers as the research object.Then,the paper analyzes the impacts of seven different preprocessed methods on the drug NIR data by constructing different layers of LSTM,BiLSTM and GRU networks and compares different classification model indicators and training time of each model.When the spectrum data are pre-processed by z-score normalization,the GRU-3 model has the best accuracy in all models.The BiLSTM models are better for analyzing high coincidence data.The method proposed in this paper can be further extended to other NIR spectroscopy data sets.

Automatic Product Image Classification with Multiple Support Vector Machine Classifiers

For the task of visual-based automatic product image classification for e-commerce,this paper constructs a set of support vector machine(SVM) classifiers with different model representations.Each base SVM classifier is trained with either different types of features or different spatial levels.The probability outputs of these SVM classifiers are concatenated into feature vectors for training another SVM classifier with a Gaussian radial basis function(RBF) kernel.This scheme achieves state-of-the-art average accuracy of 86.9%for product image classification on the public product dataset PI 100.

A MULTIPLE INTELLIGENT AGENT SYSTEM FOR CREDIT RISK PREDICTION VIA AN OPTIMIZATION OF LOCALIZED GENERALIZATION ERROR WITH DIVERSITY

Company bankruptcies cost billions of dollars in losses to banks each year. Thus credit risk prediction is a critical part of a bank＇s loan approval decision process. Traditional financial models for credit risk prediction are no longer adequate for describing today＇s complex relationship between the financial health and potential bankruptcy of a company. In this work, a multiple classifier system （embedded in a multiple intelligent agent system） is proposed to predict the financial health of a company. In our model, each individual agent （classifier） makes a prediction on the likelihood of credit risk based on only partial information of the company. Each of the agents is an expert, but has limited knowledge （represented by features） about the company. The decisions of all agents are combined together to form a final credit risk prediction. Experiments show that our model out-performs other existing methods using the benchmarking Compustat American Corporations dataset.

Research on Remote Sensing Image of Land Cover Classification Based on Multiple Classifier Combination

This paper proposed an algorithm in which the maximum probability and the weighted average strategy were used for the combination of member classifiers. Using parallel computing, we test the algorithm on a China-Brazil Earth Resources Satellite （CBERS） image for land cover classification. The results show that using three computers in parallel can reduce the classification time by 30%, as compared with using only one computer with a dual core processor. The accuracy of the final image is 93.34%, and Kappa is 0.92. Multiple classifier combination can enhance the precision of the image classification, and parallel computing can increase the speed of calculation so that it becomes possible to process remote sensing images with high efficiency and accuracy.

Multi-Domain Sentiment Classification with Classifier Combination

State-of-the-arts studies on sentiment classification are typically domain-dependent and domain-restricted. In this paper, we aim to reduce domain dependency and improve overall performance simultaneously by proposing an efficient multi-domain sentiment classification algorithm. Our method employs the approach of multiple classifier combination. In this approach, we first train single domain classifiers separately with domain specific data, and then combine the classifiers for the final decision. Our experiments show that this approach performs much better than both single domain classification approach （using the training data individually） and mixed domain classification approach （simply combining all the training data）. In particular, classifier combination with weighted sum rule obtains an average error reduction of 27.6% over single domain classification.

A fast automatic recognition and location algorithm for fetal genital organs in ultrasound images

Severe sex ratio imbalance at birth is now becoming an important issue in several Asian countries. Its leading immediate cause is prenatal sex-selective abortion following illegal sex identification by ultrasound scanning. In this paper, a fast automatic recognition and location algorithm for fetal genital organs is proposed as an effective method to help prevent ultrasound technicians from unethically and illegally identifying the sex of the fetus. This automatic recognition algorithm can be divided into two stages. In the 'rough' stage, a few pixels in the image, which are likely to represent the genital organs, are automatically chosen as points of interest (POIs) according to certain salient characteristics of fetal genital organs. In the 'fine' stage, a specifically supervised learning framework, which fuses an effective feature data preprocessing mechanism into the multiple classifier architecture, is applied to every POI. The basic classifiers in the framework are selected from three widely used classifiers: radial basis function network, backpropagation network, and support vector machine. The classification results of all the POIs are then synthesized to determine whether the fetal genital organ is present in the image, and to locate the genital organ within the positive image. Experiments were designed and carried out based on an image dataset comprising 658 positive images (images with fetal genital organs) and 500 negative images (images without fetal genital organs). The experimental results showed true positive (TP) and true negative (TN) results from 80.5% (265 from 329) and 83.0% (415 from 500) of samples, respectively. The average computation time was 453 ms per image.