Information technology has fundamentally changed different cultural groups and social structure of modern society, influenced interaction patterns and relations between individuals and organizations, and increased the complexity and interaction intensity of society. A lot of new social phenomenon and social problems have emerged quickly and traditional approaches cannot analyze these problems effectively. As a new research field of using computational techniques to tackle social problems, social computing has increasingly gained attention in recent years. Groups are the main constituent of the society and their behavior are the main focus of social computing research. Modeling and forecasting group behavior with computational approaches can help discover and understand the underlying mechanism. However, it is rather challenging to analyze group behavior. It is partly due to the high complexity of group behavior. Another main reason is the lack of historical data. This poses great difficulties for computational modeling and experimental validation and thus hinders the computational studies of group behavior. To address these challenges, we investigate two types of modeling and forecasting approaches from a social computing perspective, i.e., data-driven modeling and forecasting and knowledge-driven modeling and forecasting. The contributions are listed as follows. First, we investigate data-driven modeling and forecasting approaches for group behavior. We choose social and cultural modeling as a representive approach. Based on the benchmark datasets of group behavior, we conduct experiments to compare and evaluate the performance of social and cultural modeling and other six classical forecasting approaches with accuracy, recall, AUC measures. We discover that all these algorithms achieve high accuracy and low recall. We point out the class imbalance problem in data-driven modeling and forecasting and propose possible solutions. We further discuss the strenghs and weaknesses of data-driven modeling and forecasting approaches. Second, to acquire group behavior datasets, we propose an approach to extract the knowledge of group actions from online news and construct the plans of groups automatically. We first define the standard form of group actions and states and design an information extraction approach to extract group actions and associated states (i.e., preconditions and effects). To complement the missing action knowledge, we further refine the extr...
修改评论