With the rapid development of computer and artificial intelligence, image-based methods for skin analysis have achieved preferable results. However, the performance of computer aided diagnosis systems based on deep learning methods relies on big medical data labeled by domain experts. In addition, there is limitation of interpretability for the diagnosis results. To address aforementioned problems, a vision-based unified framework for dermatological analysis termed as parallel skin is proposed. Inspired by the ACP method and the parallel medical image analysis framework, the artificial skin image system to perform data selection and generation is constructed. Then, computational experiments are conducted with predictive learning for model building and evaluation. Descriptive and prescriptive learning to leverage the power of domain knowledge to guide data selection and generation are further introduced. In the proposed parallel-skin framework, the closed-loop diagnostic analysis model can be optimized.

The existing algorithms cannot produce satisfactory results with both low calculation cost and good fitting effect, due to the regression problems based on the complex data with large scale and non-stationary variation in industry, information and other fields. Therefore, a distributed regularized regression learning algorithm based on multi-scale Gaussian kernels is proposed. The hypothesis space of the proposed algorithm is a sum space composed of reproducing kernel Hilbert spaces generated by multiple Gaussian kernels with different scales. Since each disjoint subset partitioned from the whole data set with different degree of fluctuation, kernel function approximation models with different combination coefficients are established. According to the least square regularized method, a local estimator is learned from each subset independently in the meantime. Finally, a global approximation model is obtained by weighting all the local estimators. The experimental results on two simulation datasets and four real datasets show that the proposed algorithm reduces the running time successfully with a strong fitting ability compared with the existing algorithms.

The definition of n-step boundary set is given based on rough approximation operator in boundary region. The definition of n-step rough approximation operator is introduced and a set of stepped approximation method of rough set theory is constructed. The examples and related proofs show that positive integers always exist in both binary relations and coverage environments, the n-step upper and lower approximation sets are exactly equal to the set of objects, and the object set is the exact set in this sense. Its n-step upper and lower approximation sets or approach a fixed set of objects. The n-step rough set can make an object set approach itself or a fixed set.

Collaborative filtering(CF) cannot provide personalized recommendation with both good accuracy and diversity. To address this problem, a covering reduction collaborative filtering(CRCF) is proposed in this paper. The covering reduction algorithm in covering based rough sets is combined with user reduction in CF, and redundant elements of covering are matched with redundant users of a neighbor. The redundant users are removed by covering reduction algorithm to ensure high effectiveness of the neighbor of a target user in CF. Experimental results on public datasets indicate that CRCF provides personalized recommendations for target users with both satisfactory accuracy and diversity in sparse data environment.

The quantum state estimation and filtering problems under different interference and noise conditions are studied, and the corresponding efficient convex algorithms for quantum state density matrix reconstruction are proposed. A quantum state filtering algorithm is proposed for the simultaneous existence of sparse state interference and measurement noise. Two different quantum state estimation algorithms for the existence of sparse state interference and measurement noise are proposed, respectively. In the simulation experiments of the 5-qubit density matrix estimation, the performances of the three proposed algorithms are compared, and the performances of the three algorithms with different measurement rates are analyzed. Experimental results show that the three algorithms have high convergent speeds and low computational complexity, low estimation errors and measurement rates.

To improve classification accuracies of hyperspectral remote sensing images and make full use of local information, a feature extraction method for hyperspectral remote sensing images based on local Fisher discriminant analysis with wavelet kernel is proposed. Wavelet kernel function is introduced to map data from a low dimensional space to a high dimensional feature space, and a weighted matrix is employed to calculate scatter matrices. Local Fisher discriminant criterion function is solved to obtain the optimal feature matrix and a better separation in high-dimensional feature space. Experimental results on two open hyperspectral datasets show that the overall classification accuracy and Kappa coefficient of the proposed method are improved compared with other methods.

To take full advantage of inter-band correlations of multispectral remote sensing images, a multispectral remote sensing image segmentation method based on Gaussian copula function is proposed. Firstly, the Markov random field model is exploited to establish a label field and the label field is characterized by the Potts model. Then, the feature field characterizing pixel spectral measurements is built. A multivariate statistical model based on Gaussian copula modeling pixel spectral measurement is proposed. Furthermore, a posterior probability model of multispectral remote sensing image segmentation is established by Bayes theorem combined with the label field model, the feature field model and the prior probabilities of model parameters. The Metropolis-Hastings algorithm is designed to simulate the posterior probability model, and the optimal segmentation is obtained under the maximum a posterior strategy. Experiments are carried out with simulated and real multispectral images respectively, and experimental results indicate that the proposed algorithm has a strong ability to describe the correlation between bands with a high accuracy.

The entity linking task of knowledge base question answering(KBQA) is to accurately link the content of questions to the entities in the knowledge base. Recall rate and accuracy of linked entities cannot be balanced by most of the current methods, and only the text information is applied to distinguish and filter the entities. Therefore, a multi-dimensional information integration based entity linking for KBQA(MDIIEL) combining multi-dimensional features based on merging substeps is proposed in this paper. By representing learning methods, information such as text symbols, entities and question types, and semantic structure expressions of entities in the knowledge base are integrated and introduced into the entity linking task. The differentiation of similar entities is strengthened, and candidate sets are reduced while the accuracy is improved. The experiment proves that the MDIIEL model makes a holistic improvement on the entity linking task compared with the current methods, and it achieves the best current linking results on most indicators.

Most subspace clustering methods cannot capture geometric structures of data effetively while mapping high-dimensional data into a low-dimensional subspace. Aiming at this problem, a deep subspace clustering algorithm with low rank constrained prior(DSC-LRC) is proposed, maintaining both global and local structure information. Low-rank representation(LRR) is combined with depth autoencoder, global structures of data are captured by low rank constraint, and potential characteristics of constrained neural network are represented as low rank. Data are nonlinearly mapped into a latent space by minimizing differences between reconstructions and inputs with the local features of the data maintained. Multivariate logistic regression function is considered as a discriminant model to predict subspace segmentation. Parameters updating and clustering performance optimization are conducted in an unsupervised joint learning framework. Experiments on five datasets validate the effectiveness of DSC-LRC.

An end-to-end neural user preference modeling framework incorporating knowledge graph into recommender systems, neural user preference modeling framework based on knowledge graph(NUPM), is proposed aiming at the limitations of the current feature-based and path-based knowledge aware recommendation method. Historical interaction items of users in knowledge graph are considered as preference origin of NUPM. Then, potential preferences of users are learned by propagating user interests through relational links between entities in knowledge graph. Furthermore, an attention network is exploited to combine the preference features of different propagation stages to construct final user preference vector. The experimental results on real dataset show the effectiveness of NUPM in personalized recommendation for characterizing user preference.

To avoid the difficulties in choosing and explaining the visual features and pooling strategies in image quality assessment, a reference images quality assessment method based on multi-feature fusion is proposed. Various underlying features of reference and distorted images are extracted,and a machine learning method is applied to predict the quality of real images. Firstly, phase congruency, gradient, visual saliency and contrast of reference and distorted images are extracted. Then similarity maps of four features are calculated, respectively. The mean and variance characteristics of these similarity maps are extracted. Finally, the assessment model is learned by support vector regression. The experimental results on four benchmark databases demonstrate a high coherence between subjective and objective assessment by the proposed method.