Information Recommendation Algorithms Based On Individual Behavioral Characteristics On Social Networks

The advent of rapid development of Internet and ubiquity of electronic devices has brought unprecedented changes to daily lives.Nowadays,people can freely access to abundant online information and resources without the limitations of time and space.While the Internet makes our lives more convenient,it brings a serious problem of information overload.With applications been widely used in many aspects of our lives,recommendation systems have been thought to be the most effective tools to deal with the information overload problem.Recommendation systems not only help customers to dig out relevant information and potential interested articles and services,but also contribute to increase the visibility of products and maintain the loyalty of customers.In recent decades,many researches have been reported in the field of recommendation systems.It becomes a hot issue that people from a variety of disciplines such as economics,computers and social sciences have paid great attentions to recently.Recent frontier theoretical studies have greatly promoted the innovativeness and real-world applications of recommendation systems.However,there are also some remained difficulties and challenges for the current studies.For instance,the apparent diversity-accuracy dilemma of recommendation systems has long been a main issue.Too much attention to the recommendation accuracy would decrease the diversity of recommended objects and narrow down the options of customers because of the larger similarity of recommended objects in the future.Furthermore,cold-start is another major problem faced by recommendation systems.The ability of recommendation system to deal with new coming users and products is weakened by the lack of corresponding review information in the system.Under this condition,the system would provide unreasonable recommendations.There are two potential solutions to these problems.On the one hand,since most of the recommendation systems are based on the calculation of similarities among users or objects,it is promising to balance accuracy and diversity by designing more reasonable similarity index.On the other hand,in terms of the cold-start problem,the performance of recommendation systems could be improved by introducing additional information such as the trust relations between users and the interest tags of users into the design of recommendation algorithms.This thesis focuses on the diversity-accuracy dilemma and the cold-start problem.On addressing the diversity-accuracy dilemma,how vertex similarity for bipartite networks affect the recommendation performance is analyzed,and a novel vertex similarity index is proposed.The vertex similarity-based recommendation method improves both the recommendation accuracy and diversity.Then,by introducing additional trust relations information between users,two trust enhancement hybrid recommendation methods are proposed respectively under the similarity-based framework and resource-allocation framework.With the help of the additional information,both accuracy and diversity are significantly improved.On addressing the cold-start problem,user preferences to objects are calculated under the framework of the gravity model by taking into account the information of the user interest tags.The tag-based recommendation method not only improves the performance of recommendation system but also solves the cold-start problem.Meanwhile,we conducted further analyses on label information from two perspectives:the evolution of label-based recommendation networks,and the users' adoption of label information.The detailed analyses and the key points are summarized as follows:1.Studies on the similarity-based recommendation method.For similarity-based recommendation methods,we explored how the selection of similarity index affects the accuracy,diversity,and novelty of recommendation results.Results showed that cosine similarity based recommendation methods bring high accuracy but relative low diversity.By comparison,resource-allocation based recommendation methods advantage diversity but weaken accuracy.Combing both the cosine similarity index and the resourceallocation index,we proposed a novel vertex similarity index,named CosRA index.We applied the CosRA into recommendation systems under the resource-allocation framework.Experimental results based on several real-world datasets showed that the CosRAbased recommendation method improves accuracy,diversity,and novelty at the same time.The CosRA index is parameter independent,and we found generalized CosRA index could not remarkably improve the recommendation performance,suggesting that the current form of the CosRA index is already optimal.According to our results obtained from empirical dataset,the proposed CosRA index has a significant advantage in applications of recommendation systems.2.Studies on the trust relationship enhancement recommendation method.By using the trust relationships between users in the recommendation systems,we proposed the trust enhancement recommendation method.Specifically,in the resource-allocation process of the CosRA-based method,the amount of resources received by the trusted users is adjusted by a tunable parameter before allocating to users.Experimental results suggested that there is an optimal value of the parameter for the higher accuracy,and the optimal parameter value is nearly identical among several datasets under different evaluation metrics.Considering that the tunable parameter controls whether the amount of resources users is enhanced or suppressed,our results showed that moderate amount of trust information could improve the performance of system while over amount of information would impair the performance.Furthermore,we applied the same idea under the framework of the resource-allocation recommendation systems.We proposed the trust enhancement resource-allocation recommendation method and analyzed how the amount of the trust relationships affect the recommendation performance.Experimental results suggested that enhancing the weight of trust users on reallocating resources can improve the performance.3.Studies on the tag-based recommendation method.To address the cold-start problem,user interest tags are introduced as another complementary information.Interests tags refer to the information manifest user interests and preferences.We proposed an algorithm to predict the products which users may have interest on based on the tag information under the framework of the gravity model.Experimental results suggested that our method increased both accuracy and diversity compared to some benchmark recommendation methods.When recommending new products,the time complexity of the proposed tag-based recommendation method remains nearly constant.The time complexity advantage indicates our method has potential application in the synchronized recommendation system.Next,by combining the information of the user interest tags and the gravity model,we analyzed the evolution of the user attributes and the object properties under several classical network models.Further,we proposed a non-Markovian model to describe information spreading on multiplex social network by taking the utilization of tag information and information spreading patterns into consideration.Our model contributes to explain sophisticated social information spreading phenomenon.