Topic Network Analysis Based on Co-Occurrence Time Series Clustering

Abstract

Traditional topic research divides similar topics into the same cluster according to clustering or classification from the perspective of users, which ignores the deep relationship within and between topics. In this paper, topic analysis is achieved from the perspective of the topic network. Based on the initial core topics obtained by the keyword importance and affinity propagation clustering, co-occurrence time series between topics are constructed according to time sequence and topic frequency. Subsequence segments of each topic co-occurrence time series are divided by sliding windows, and the similarity between subsequence segments is calculated. Based on the topic similarity matrix, the topic network is constructed. The topic network is divided according to the community detection algorithm, which realizes the topic re-clustering and reveals the deep relationship between topics in fine-grained. The results show there is no relationship between topic center representation and keyword popularity, and topics with a wide range of concepts are more likely to become topic network centers. The proposed approach takes into account the influence of time factors on topic analysis, which not only expands the analysis in the field of topic research but also improves the quality of topic research.

Unleashing Competitive Intelligence: News Mining Analysis on Technology Trends and Digital Health Driving Healthcare Innovation

“…First, the entities are clusterized as described in Table I, analogously as in graph-based topic modeling techniques [83], [92], [93]. We observe that the nodes are grouped into six main clusters, representing 97.52% of the total number of entities in the network.…”