Extracting actionable knowledge from social networks with node attributes. (September 2019)

• Record Type: Journal Article
• Authors: Kalanat, Nasrin; Khanjari, Eynollah
• Is Part Of: Expert systems with applications. Volume 3(2019)
• Publication Date: 2019-09
• Subjects: Social network -- Node attribute -- Actionable knowledge discovery -- Action extraction -- Random walk; 006.33
• DOI: 10.1016/j.eswax.2019.100013
• Languages: English
• ISSNs: 2590-1885

Abstract

Highlights: Introducing the extraction of actionable knowledge from social networks involving node attributes. Formulating the action extraction process as an optimization problem. Extracting mostly qualitative actions in terms of cost and effectiveness. Exploiting heuristics to extract actions more efficiently. Abstract: Actionable Knowledge Discovery has attracted much interest lately. It is almost a new paradigm shift toward mining more usable and more applicable knowledge in each specific domain. An action is a new tool in this research area that suggests some changes to the user to gain a profit in his/her domain. Currently, most of action mining methods rely on simple data which describes each object independently. Since social data has more complex structure due to the relationships between individuals, a major problem is that such structural information is not taken into account in the action mining process. This leads to miss some useful knowledge and profitable actions. Consequently, more effective methods are needed for mining actions. The main focus of this work is to extract cost-effective actions from social networks in which nodes have attributes. The actions suggest optimal changes in nodes' attributes that are likely to result in changing labels of users to more desired one when they are applied. We develop an action mining method based on Random Walks that naturally combines the information from the network structure with nodes attributes. We formulate action mining as an optimization problem where the goal is to learn a function that varies the values of nodes' attributes which in turn affect edges' weights in the network so that the labels of intended individuals are likely to take the desired label while minimizing the cost of incurring the changes. Experiments confirm that the proposed approach outperforms the current state-of-the-art in action mining.