Temporal Topic Correlation and Evolution

Abstract

This paper presents a technique to alternatively discover the temporal research topics correlation by using a topic model, Latent Dirichlet Allocation (LDA). LDA model assumes the documents as a mixture of topics that group the co-occurrence words with the certain probabilities. Hence the model is popularly used to extract the latent topics from document collections. However, LDA gives an independence assumption between topics and is unable to model the correlation between the topics. Motivated by above limitation, this study introduces a method for improving the topic correlation. The correlation of two topics from different time periods can occur when there exists a publication tagged by the two topics and these two topics are said to be co-occurred by this publication. LDA weights of these co-occurred topics are used in our model to calculate gross-correlation values. The number of publications in a topic co-occurrence is also used in the model. Therefore, we split dataset into groups with some common sub-dataset ordered by temporal timestamp of published year. The experiment results show the correlation between topics in different time periods and results can further support the research collaboration in future.