Multiple contradictory attribute information makes it difficult for decision makers to make decisions in multi-attribute decision-making, and therefore the problem are studied from the perspective of support intuitionistic fuzzy sets in this paper. Firstly, on the basis of support intuitionistic fuzzy sets, two models of optimistic and pessimistic multi-granulation support intuitionistic fuzzy rough sets are constructed in combination with the theory of multi-granulation rough sets. The relationship between two models above is analyzed and the related properties are discussed. Then, the fitting function is defined by t-norm and t-conorm, and a multi-attribute decision-making solving method with multi-granulation support intuitionistic fuzzy rough sets is proposed. Meanwhile, a score function and a accuracy function are defined to sort decision results, the corresponding decision rules are extracted, and an algorithm is designed. Example analysis verifies that the method enables decision makers to select the optimal decision-making scheme according to actual demands while dealing with conflicting multi-attribute decision-making problems.

Single model with single granularity is employed to process multi-sources heterogeneous raw data in the existing user portrait models. The performance of the analytic model is limited and the multi-level and multi-angle user portrait features cannot be fully displayed. Aiming at this problem, based on the idea of granular computing, a multi-granularity user portrait model is proposed. Firstly, a multi-granular representation structure of the data is constructed to granulate the raw data. Then, according to the data granularity structure, a granularity upgrade algorithm based on ensemble learning is proposed. Low-level data information is fused to obtain high-level data representation. Finally, user portrait analysis is carried out at multi-level data representation to show a more comprehensive portrait. Experiments show that the user portrait with multiple granularities is more comprehensive, stereoscopic and richer than the single granularity user portrait.

It is difficult to calculate attribute reduct due to high time and space complexity of entire-granulation rough sets. To solve this problem, distinguishability in information systems is defined by equivalent class, and its properties are studied. It is proved that the attribute reduct based on distinguishability is equivalent to absolute reduct. The positive region distinguishability in decision systems is defined and its properties are discussed. It is also proved that positive region distinguishability reduct is a superset of entire-granulation Pawlak reduct, but in most cases it is equal to entire-granulation Pawlak reduct and it can be regarded as an approximation of entire-granulation Pawlak reduct. Theoretical analysis and experiments show that compared with other attribute reduction algorithms, positive region distinguishability reduct has great advantages in computational complexity and classification accuracy.

The existing multi-label feature selection algorithms based on fuzzy rough sets characterize the ability of distinguishing attributes from single sample space, while the ability of attributes distinguishing labels is ignored. Therefore, a multi-label feature selection algorithm based on fuzzy discernibility relations in double spaces is proposed. Firstly, two multi-label attribute measures based on fuzzy discernibility relations are defined from the perspective of samples and labels respectively. Then, two different measures are combined by introducing weights. Finally, a multi-label feature selection algorithm is constructed based on the combined measures by utilizing the forward greedy algorithm. Results of comparative experiments on five multi-label datasets verify the effectiveness of the proposed algorithm.

In multi-granulation rough set models, granulation selection is always related to positive region. Due to the excessive classification on the object set determined by all labels, few or none objects fall into the positive region, and a lot of information may be lost or even fail in positive reduction methods. To overcome this deficiency, an algorithm of optimal granulation selection for multi-label data based on local generalized multi-granulation rough set is proposed. Firstly, local generalized multi-granulation rough set model is introduced in multi-granulation and multi-label information system. Information level parameters are set, and the target set according to each label is approximated. The granularity quality of the multi-granulation and multi-label information system is defined, and then granular significance is obtained. Finally, a heuristic algorithm for optimal granularity selection is designed, and its effectiveness is verified.

In many real world applications, data is dynamically generated at different time periods in addition to high-dimensional imbalanced features. An high-dimensional class-imbalanced online feature selection algorithm based on neighborhood rough set is proposed. The algorithm design is based on rough dependency calculation formula of small class significance. Meanwhile, three evaluation criteria of online relevance analysis, online redundancy analysis and online significance analysis, are presented to select features with high separability between majority and minority classes. Experimental results on seven high-dimensional and class-imbalanced datasets show that the proposed method can effectively select a better feature subset with better performance.

Time complexity of the existing supervised topic model training algorithms is generally linear to the number of topics and therefore their large-scale application is limited. To solve this problem, an efficient learning algorithm for maximum entropy discrimination of latent Dirichlet allocation(MedLDA) supervised subject model is proposed in this paper. The proposed algorithm is based on coordinate descent, and the number of iterations of training classifiers is less than that of the existing Monte Carlo algorithm for MedLDA. The algorithm also makes use of rejection sampling and efficient preprocessing technique to reduce the time complexity of training from linear to sub-linear with respect to the number of topics. The comparison experiments on multiple text corpora show that the proposed algorithm makes a great improvement in training speed compared with the existing Monte Carlo algorithm.

The primary objective of a personalized recommendation system is to provide pertinent recommendations for users and improve internet information utilization. A social network personalized recommendation algorithm based on heat diffusion influence propagation(HDIP) is proposed in this paper combining HDIP with the hidden follow-up relationship in the social network of users. Firstly, in the HDIP algorithm, the friendship in real life is transformed into follow-up relationship between customers in shopping network. Heterogeneous information network graphs are constructed and the composite similarities between users are calculated. Secondly, the influence propagation process in social networks based on the heat diffusion model is simulated. Probability scores of users in the social network are calculated and accurately sorted to select neighboring users similar to the target users. Finally, the products of potential interest are recommended to the target users according to the ranking. Thus, the personalized recommendation is implemented. The public dataset is utilized for the comparison between HDIP and conventional recommendation algorithms. The experimental results show that HDIP produces a relatively high accuracy and various recommendation effects.

Most of the existing track allocation works neglect the local nets problem, and are prone to fall into the local extremums. Based on discrete particle swarm optimization, genetic operation and negotiation-based refining strategy, a track assignment algorithm is proposed by considering local nets, overlapping conflict, wirelength and blockages. The algorithm abstracts local nets and constructs the corresponding model of segments. To expand population diversity, hybrid genetic operation is incorporated to improve the efficiency of global search. At the same time, a simple and efficient fitness function is designed. Finally, the negotiation-based refining strategy is exploited to further reduce the overlap of segments. The experimental results indicate the effectiveness of the proposed algorithm. The algorithm can obtain better overlapping cost index optimization value and reduce the congestion in the key routing area.