Research On Key Issues Of Context-aware Recommender System

The era of big data has brought new development opportunities and challenge to content services under various network environments.On the one hand,it enriches the content of the service and offers more comprehensive data support.On the other hand,it is harder to excavate valuable resources for users,because the characteristics of which are high-dimensional,sparsity and low-value density,it exacerbates the influence of information overload on the acceptability of content services.As one of the core technology to solve the above problems,context-aware recommender systems have received extensive attention and further research in industrial and academic circles in recently years.But there are still exist some drawbacks for the research of support data and recommendation algorithm,which including the inaccuracy and sparsity of support data,the insufficient research of the facters that influence users' preference.Then the research of context-aware recommender systems which can solve the above problems can provide powerful promotion of information content service under big data,and has important research significance and practical value.This thesis proposed new optimizing strategy and implement method for the support data and core algorithm,and solved many key problems of context-aware recommender systems,including how to correct the fake preference behavior in users' historical preference data,how to solve the sparsity problem of users' historical preference data,how to dig the relevant relationship among users and items,how to build users' preference model with multi-view,how to use users' cognitive behavior to build preference model,and how to use clustering cognitive behavior to optimize users' preference model.The specific tasks of study are as follows:(1)To solve the problem that the existing recommendation algorithms have not modified users'historical data which don't match their preferences model,and lead to low predictive accuracy,an approach based on decision cheat adaptive modified inspired by the theory of decision-making psychology is proposed.This method first gets users' fake preference behaviors by the behavior distance of central point,then it cross validates the context behavior fluctuation and preference fluctuation,and combines the context membership function to connect the fake preference behaviors and the set of fuzzy decision cheat.Finally,it uses different behavioral compensation to elicit user preferences.The authors present empirical experiments by using a real extensive data set,the results of which show that the proposed algorithms can achieve better predictions compared with collaborative filtering algorithm and context-aware algorithm.It is been proved the ability that can effectively alleviate the influence of the fake preference behavior on recommendation quality.(2)The preference behavior similarity measure methods in existing collaborative filtering based recommender systems often fail to acquire the real nearest neighbors,which influence the prediction accuracy.A users' preference prediction method based on the optimization of basic similarity space distribution is proposed.In the beginning,this method uses measure methods to get the original similarity among users.It then computes the preference center based on preference similarity,and gets the average similarity range based on the distance among other similarity and preference center to build the basic similarity space.The method builds the modified model based on average nearest neighbor and abnormal rating to optimize basic similarity space and generate prediction.The simulation experiment results show that the proposed method can achieve more accurate nearest neighbor,and significantly improve the prediction accuracy,which indicated its ability to solve the sparsity problem.(3)To solve the problem that the existing methods only acquire users' preference from the single view which leads to low predictive accuracy and applicability,a multi-view trust relationship based context-aware method is proposed.This method first gets the trust relationship among users by direct similarity and indirect similarity,then obtains the relationship between users and items by conception similarity.Fisher linear discriminant analysis is adopted to fuse the sample data got from the two views.Then it uses projection transformation and Fisher discriminant criterion to projection direction which has the minimum within-class scatter and maximum between-class scatter.Besides,it has the best recommendation accuracy,and based on which the recommendation is generated.Experimental results show that the proposed algorithms can generate more accurate relationship among user' preference and improve the quality of recommendation quality.(4)To solve the problem that the lack of the consideration of internal relation between user's cognitive behaviors and preferences leads to low predictive accuracy in existing recommender systems,a cognitive behavior based approach is proposed.First,the paper introduces many cognition concepts into the process of acquiring user preferences under multidimensional context environment,such as cognitive level,cognitive risk,effective cognitive and so on.And then this paper provides the definitions and calculation methods of those concepts.Finally it elicits user preferences under unidimensional and multidimensional context environments by establishing the mutual effect model of cognitive factors.The empirical experiments shows that it can achieve better recommendation quality than the methods based on users' behavior,which indicates that the method can effectively reflect the relationship between user's cognitive behavior and preference,and it is verified the validity and feasibility to build the preference model based on cognition.(5)To solve the problem that the existing methods acquire users' preference only from the single view without considering the internal relation between users' cognitive psychology and preference behavior,which leads to the low predictive quality and recommendation list's failing to cover users' latent preference,a multi-view cluster cognition optimization model based context-aware recommendation method inspired by distributed cognitive model and multi-view learning is proposed.This method measures the incidence relation among users'cognitive behavior from the view of social cognition based on social networks and individual cognitive behavior,and builds the cognitive relation model which includes reliant cognition,cognition risk and cluster cognition.Then the method uses feature fusion,projection transformation and Fisher discriminant criterion to build multi-view optimal preference elicit model from the views of users' properties,cognitive relation and cognitive value of items.Experimental results show that comparing with existing methods,the proposed algorithm can achieve better prediction accuracy(about 20.19%)and diversity(about 44.89%).It verified the effectiveness of the proposed method,and shows users' cognitive behavior are influenced by relevant user group,the cluster cognitive model based on real social environment can more accurate reflect users' cognitive state,than based on single user' cognitive behavior.