Research Of Terrorist Behavior Prediction Algorithm Based On Transfer Learning

With the development of cultural modeling, the context of organization has been used to predict terrorist behavior. Recently, the research of predicting terrorist behavior based on context knowledge tends to choose the organization that contains rich samples as the prediction object. However, due to the globalization of terrorism,many new organizations are emerging, and short living time of organizations lead to the lack of available samples. For the above situation, it is difficult to effectively model the behavior of new organization by traditional method, which will increase the difficulty of behavior prediction and reduce the accuracy of prediction.Therefore, this paper makes analysis and research on the characteristics of terrorist organization datasets, and new methods are proposed as follow: 1) Terrorist behavior prediction algorithm based on multi source instance transfer learning is proposed to solved the impact of the lack of available samples on the prediction of the new organization. 2) In the course of instance transfer, due to the reducing of prediction effect caused by invalid source organization selection, Source organization selection algorithm based on attribute partition and clustering is proposed to improve the effect of prediction. The main contents of this paper are as follows:(1) This paper proposes Multi Organization Transfer Adaptive Boost Algorithm(MO-TrAdaBoost) for behavior prediction in new terrorist organization. According to the characteristics of transfer learning that can effectively solves the learning of new domain, the algorithm uses the source domain samples to assist the target organization training and iteratively correct the weight of the error sample, so as to reduce the impact of the lack of samples on the prediction. Furthermore, multi-source easily degenerates into a single source in instance transfer, which will leads to unstable prediction. Therefore, the algorithm uses the filtering mechanism and the source organization weight factor to optimize the process of knowledge transfer, which makes most of source organizations can participate in knowledge transfer and reduces the impact of organization with poor historical performance. Experiments on MAROB show that MO-TrAdaBoost improves CONVEX and SVM by at least 13.2% and 7.8for accuracy and recall, and the proposed algorithm can be well adapted to prediction of the new organization.(2) This paper proposes Source organization selection algorithm based on attribute partition and clustering to improve the effect of behavior prediction. The algorithm uses entropy to partition the attribute, which extracts the normal attributes(value is not easy to change) to reflect the inherent characteristics of the organization.Aiming at the problem of invalid source organization selection, the algorithm uses K-modes to cluster organizations based on normal attribute, which considers mining the relationship of organizations to select the matched source organization. Therefore,the algorithm solves the problem of invalid source organization selection, and due to only part of source organizations participate in iteratively training, the time cost of the improved prediction algorithm(IMO-TrAdaBoost) is reduced. Furthermore, training only based on abnormal attributes(value change frequently), which avoid the impact of the difference in organizations’ normal attributes. Experiments on MAROB show that IMO-TrAdaBoost improves MO-TrAdaBoost by at least 1.3% and 1.9% for accuracy and recall, and reduces the time overhead by 70.6%.(3) In order to verify the feasibility of the above algorithm in practical application, this paper use Java with Eclipse to design and implement the prototype system based on object-oriented theory.