Search

Least squares support vector machines (LS-SVM), with excellent generalization performance and low computational cost, has been proven to be a useful tool in consumer credit risk assessment. It is a common assumption that the labels of the consumers are unchanged, which is contradictory with population drift. In this paper, we use a fuzzy membership of each input data to represent the impact of population drift on consumers' labels and the relative importance for the construction of the separating decision function, which is an ensemble of fuzzy sets and sparse LS-SVM. The purpose is to try to explain why an applicant should be rejected. Two UCI and an American credit card datasets are used to test the efficiency of our method and the result proves to be a satisfactory one. [ABSTRACT FROM PUBLISHER]

For least squares support vector machine (LS-SVM) classifier to the loss of sparseness and generalization, a pruning modeling method is proposed based on Quadratic Renyi entropy. The kernel principal component is adopted for data pre-processing, and the training set is divided randomly. Then the concept of quadratic Renyi entropy is introduced as the basis of training and pruning in LS-SVM classifier. UCI typical datasets of classification are used for testing the performance of this new model. Experimental results show that the new algorithm takes full account the location of the Lagrange multiplier, thus the sparseness and generalization ability of the classifier can be improved. [ABSTRACT FROM PUBLISHER]

For the loss of sparseness in least squares support vector machine (LS-SVM) model, a new pruning algorithm using Renyi entropy for LS-SVM modeling is presented. The kernel principal component is adopted for data pre-processing, then the training subsets are divided randomly. To solve the problem that the conventional pruning algorithm cannot take full account the location of the Lagrange multiplier, the concept of quadratic Renyi entropy is introduced as the basis of training and pruning in LS-SVM modeling. The results of simulation verify the validity of the algorithms, thus the sparseness and generalization ability of the model can be improved. The presented algorithm can be applied to multiple-output modeling. [ABSTRACT FROM PUBLISHER]

It's worth to investigate that the false alarm is reduced with polarimetic information while keeping the detection ratio when the fully polarimetic SAR images used for ship detection. An improved S-NMF (Sparseness-Nonnegative Matrix Factorization) for ship detection according to the nonnegative and sparseness characteristics of eigenvalues is proposed in this paper. The first two eigenvalues are used for NMF decomposition since they take most of the energy of the targets, and then we combine the advantage of the third eigenvalue which could get rid of the false alarms for ship detection with the OS-CFAR method. The proposed method is validated using Radarsat-2 fully polarimetric images with the corresponding AIS data. Results show that it cannot only detect most of the ships, but also reduce the false alarms effectively, and has good ability when the detection conditions change. [ABSTRACT FROM PUBLISHER]

Recently, we developed context-dependent deep neural network (DNN) hidden Markov models for large vocabulary speech recognition. While reducing errors by 33% compared to its discriminatively trained Gaussian-mixture counterpart on the switchboard benchmark task, DNN requires much more parameters. In this paper, we report our recent work on DNN for improved generalization, model size, and computation speed by exploiting parameter sparseness. We formulate the goal of enforcing sparseness as soft regularization and convex constraint optimization problems, and propose solutions under the stochastic gradient ascent setting. We also propose novel data structures to exploit the random sparseness patterns to reduce model size and computation time. The proposed solutions have been evaluated on the voice-search and switchboard datasets. They have decreased the number of nonzero connections to one third while reducing the error rate by 0.2–0.3% over the fully connected model on both datasets. The nonzero connections have been further reduced to only 12% and 19% on the two respective datasets without sacrificing speech recognition performance. Under these conditions we can reduce the model size to 18% and 29%, and computation to 14% and 23%, respectively, on these two datasets. [ABSTRACT FROM PUBLISHER]

When applied to speech, Non-negative Matrix Factorization is capable of learning a small vocabulary of words, foregoing any prior linguistic knowledge. This makes it adequate for small-scale speech applications where flexibility is of the utmost importance, e.g. assistive technology for the speech impaired. However, its performance depends on the way its inputs are represented. We propose the use of exemplar-based sparse representations of speech, and explore the influence of some of these representation's basic parameters, such as the total number of exemplars considered and the sparseness imposed on them. We show that the resulting learning performance compares favorably with those of previously proposed approaches. [ABSTRACT FROM PUBLISHER]

The use of multilevel modeling with data from population-based surveys is often limited by the small number of cases per macro level group, due to the fact that the data were not specifically collected hierarchically. This data sparseness problem can render multilevel modeling inappropriate because of insufficient within-group variability, and can also lead to convergence problems in more complex multilevel models. As a result, many studies that examine contextual effects with population-based survey data are forced to ignore the hierarchical structure of the data altogether by resorting to OLS regression techniques. We used a simple strategy to reduce data sparseness in a hierarchical cross-classified model using a cluster analysis procedure to group single observation neighborhoods into larger "synthetic" neighborhoods with similar characteristics. In this paper we present the details of this technique and outline its effect on our results and the relative performance of the models with and without these "synthetic" neighborhoods. [ABSTRACT FROM AUTHOR]