MDCPD:基于矩阵序列距离度量的数字生态变点检测

A digital ecosystem is a distributed, adaptive, and open social-technical system that possesses characteristics similar to natural ecosystems, such as self-organization, scalability, and sustainability. With the development of technologies such as big data, Internet of Things, and cloud computing, the concept of digital ecosystem has become increasingly complex and diverse, becoming more closely related to people's lives. Digital ecosystem evolves continuously due to the internal and external impacts. Some events may significantly change the properties of the digital ecosystem and make it deviates from its normal evolutionary path, thus causing the ecosystem unhealthily evolve with anomalies. If these changes can be detected and human intervention is carried out in time, the negative impact may be minimized. Digital ecosystem is observable through event flow, which naturally forms a network structure and makes it possible to analyze the evolution status of a digital ecosystem. Under this consideration, it is a mainstream to use the complex network to model a digital ecosystem, and change point detection in such field is one of the main techniques to detect evolution status of a digital ecosystem. However, few existing approach on change point detection are optimized for the characteristics of the digital ecosystems. They often overlook the dynamics property as they prefer discrete-time dynamic graph for modelling. This results in they are not able to support the detection of change points in a digital ecosystem under dynamic and continuous changing scenario, which brings the decrease of the performance in change point detection task for digital ecosystem. To address the problem, n this paper, we propose Change Point Detection for Digital Ecosystem Based on Sequenced Matrices Distance Measurement (MDCPD). MDCPD is a community-aware approach, which utilizes the community structure matrix sequence efficiently to assess the happened changes from community perspective. Therefore, MDCPD can detect change points in digital ecosystem under continuous-time dynamic network modelling as well as locate the reason for evolution at event-level. Based on two distance measurement methods, we develop two variants of MDCPD, and their advantages are illustrated in the paper. To reduce the noise in the matrix sequence, we propose the matrix intervention strategy. The strategy enhances the expressivity of community structure matrix sequence according to the weight matrix observed in the digital ecosystem, which improves the performance of MDCPD. We show the effectiveness of MDCPD through the comparison with the state-of-the-art (SOTA) approaches on both continuous and discrete scenarios, where each scenario contains two synthetic datasets. MDCPD exceeds the SOTA approaches by 0. 383 and 0. 034 on F1 metric on continuous and discrete scenario respectively. An ablation study is carried out on the same datasets to show the effectiveness of matix intervention strategy. The intervention strategy improves the performance of MDCPD by 0. 053 on F1 metric at most. Finally, a case study is conducted on a real digital ecosystem dataset. We demonstrate how MDCPD can be used to analyze digital ecosystem evolution and the potential conclusions that can be drawn by a combination of text and visualization. This suggests the practical value of MDCPD.