| 英文摘要 | In recent years, the number and activeness of various national and international social organizations have increased rapidly. They have had increasingly profound impact on the political, economic and other situations worldwide. Consequently, there is great demand for studying the behavioral patterns of social organizations to facilitate the decision making processes of related stakeholders, such as governments. With the rapid development of new technologies such as the Internet, information has become unprecedentedly rich and easy to access. As a result, efficient computing technologies have demonstrated great advantage on modeling organizational behavior, and become the main research trends. Thus, modeling organizational behavior has become one of the major themes in social computing. At present, research on modeling organizational behavior has focused on building predictive models using machine learning methods, in particular classification methods. In this dissertation, we first evaluate the predictive models constructed by seven representative classification algorithms and gain a thorough insight into the pros and cons of each algorithm. Secondly, we found that in organizational behavior data, class imbalance and non-uniform misclassification costs are pervasive in this domain, which severely hinder the performance of standard classifiers. To handle this problem, we empirically investigate four typical cost-sensitive learning methods, combined with six standard classifiers. Our empirical study verifies the effectiveness of cost-sensitive learning in organizational behavior modeling. Based on the experimental results, we gain a thorough insight into the problem of class imbalance and non-uniform misclassification costs, as well as the selection of cost-sensitive methods, base classifiers and method-classifier pairs for this domain. We also propose an improved algorithm which outperforms the best method-classifier pair using the benchmark organizational behavior data. In addition, we further propose a personalized solution based on cost curves. Given a dataset, this solution makes it very easy for the user to select the best method-classifier pair. Although predictive models provide accurate predictions of organizational behavior, they do not directly and explicitly suggest specific actions to take to influence (restrain or encourage) the behavior for the user’s interest. The user, however, often exactly needs such kind of actionable knowledge. Acti... |
修改评论