Computational Intelligence Models And Algorithms In Complex Information Networks

The complex information network is a kind of complex information system which contains complex data,complex information and complex knowledge.Com-pared with other kind of information agent,complex information networks have their own characteristics:the data related to each other on the network and the net-work includes much more information than we can mined.In aims of mining useful information from the complex information network and analysis the complex data structures of multi-domain data,heterogeneous data and unbalanced data across do-mains in the complex information network,we proposed a series of computational intelligence models and algorithms to provide more methodologies to applications-in the information service areas.(1)Computational intelligence models for multi-domain dynamic information transferIn the view of the dynamic information transfer,we propose a self-adaptive multi-view model for multi-source information service in intelligent transport sys-tems.Meanwhile,motivated by genetic immune process,we propose a SEIR immune strategy based instance weighted Naive Bayes classification model.The self-adaptive multi-view model for multi-source information service in intelligent transport systems is a kind of data-driven intelligent service model.It consists of a Newton iteration algorithm,a multi-layer feed forward neural network model and a infinite mixed distribution representation model.The model is experimented by intelligent transportation datasets.The SEIR immune strategy based instance weighted Naive Bayes classification model also learns by data driven and self adap-tations and shows good experimental results evaluated by classification accuracies on a standard machine learning dataset and real-world datasets.(2)The unsupervised transfer learning algorithms in computational intelligence The transfer learning framework extends the computational intelligence algo-rithms by transfer knowledge across domains.It focuses on the scenarios that the samples in the target domain are not enough for a good learning and the data structure are not the same across the source domains and the target domains.In these cases,the traditional computational intelligence algorithms cannot output a good result.We also fully consider the abundant unlabeled data in the real world,and proposed the novel unsupervised transfer learning algorithms:heterogeneous feature space based unsupervised transfer learning algorithm and multi-instance graphical transfer clustering algorithm for complex transportation service network.The former algo-rithm studies unsupervised learning and heterogeneous feature learning,improves the traditional SVM,and proposed the unsupervised multi-task selection machine for source domains.The heterogeneous feature space based task selection machine achieves the intelligent computation task in the target domain by cross-domain knowledge transfer and evidenced by experiments.The latter algorithm in this part extends multi-instance learning.It consists of an instance representation algorithm,a kernel model combination algorithm,an intelligent clustering initialization algo-rithm and an intelligent clustering update strategy algorithm.The computational performance of the proposed multi-instance graphical transfer clustering algorithm for complex transportation service network is examined by a complex traffic network dataset.(3)The computational intelligence model and algorithm for complex informa-tion networksBy comparing with the power-indicator-based node importance intelligent eval-uation model,the centrality-indicator-based node importance intelligent evaluation model,the multi-indicator-based node importance intelligent evaluation model,and the dynamic-spread-mechanism-based node importance intelligent evaluation model,we study the innovations of the computational intelligence models of the node importance evaluation in complex information networks.By considering global network connections,we propose a novel intelligent node importance evalua-tion algorithm for complex information networks,namely a global-recursive-based node importance intelligence evaluation algorithm.The algorithm take the connec-tions of all the nodes in the network into consideration,trains the global weights of the network connections to denotes the relationships between nodes,and introduces the parameter training into a Friedmann basic network evaluation.The experimental results on the Friedmann network proves the algorithm performance.(4)The data-driven multi-domain intelligent information computation:Case studiesWe study the data-driven multi-domain intelligent information computation based on specific cases.To analysis the users' behaviors in the data-driven ur-ban intelligent transportation networks,we proposed a 5S model with 5S factor analysis,model 5S structure designs and a Map-Reduce learning process design.Comparing with the samples that preprocessed by PCA dimensional reduction,the overall sample idea is videnced to be suitable for the data-driven multi-information computational intelligence cases.The main contributions are as follows:(1)the knowledge is successfully transferred by transfer learning models and algorithms on the complex information network;(2)the unsupervised learning algorithms fully mines useful information from real-world unlabeled data;(3)dynamic information networks achieve self-adaptive learning in the cross-domain information transfer;and(4)the models and algorithms which proposed based on the complex information network transfer learning well serve the real-world computational intelligence with the complex information network.