Multimedia Data Mining In Social Network Sites

In recent years, with the development of computer hardware and network infrastructure, Social network sites becoming the most popular network application. A variety of social networks gradually deeply into our lives and closely related with the users. Hence, thinking of some researchers resulting from the academic issues that brought by SNS. Researching on social network sites mainly focus on the following two issues:1. How to mining social network sites reasonably and effectively. Through mining social network, we can predict the habits, interests and hobbies of users and then conduct targeted information recommendation to enhance the user experience.2. How to manage the fast-growing social network sites efficiently. With the increasing of users and user stickiness, the multimedia documents in social network growing exponentially. Thus, it is necessary to propose some effective methods to manage massive network data, including storage and query.Social network sites have two significant advantages: high user stickiness and high accurate of real-name information. With these advantages, this paper propose some innovative research programs to solve the above two issues.For issue one, this paper add image content based data mining to social network data mining innovatively, deepen the traditional graph data mining that based on history operation records. Image content based data mining use the relationship between the characters in the image and text labels to gain the more convincing relationship judgments. There are two ways to conduct information recommendation in social network, link-type and cluster-type. Link-type recommendations use the network relationship to query the communication links and then recommend friends. Cluster-type recommendations use the attributes of users to clustering, and then send advertisements information to the centers of clustering. This paper develop solutions for both link-type and cluster-type recommendations. For issue two, this paper design and improve the massive multimedia data index based on local sensitive hash, reduce the time complexity of multimedia data retrieval. We also optimize the popular distributed systems in a targeted manner and make users retrieving multimedia data in real time come to truth.As the diversity of network data, different mining methods should be used to adapt different data structures. However, it is inevitable to gain different knowledge networks and how to ensemble these knowledge effectively is a great challenge. This paper use the semi-supervised learning algorithms which based on graph maximum confidence to ensemble the knowledge that mined from text, image and history operation records. The experiment results show that the ensemble algorithm is both effective and efficient.