Performance Evaluation of Probabilistic Latent Semantic Analysis for Unstructured Social Media Data.

摘要

Big data analytics is being applied in many fields today to mine unstructured data such as social media blogs or medical records. We focus this thesis on two popular analysis techniques, the methods of Latent Semantic Analysis(LSA) and Probabilistic Latent Semantic Analysis(PLSA), both used for interpreting or extracting concepts and relationships from data. As a use case, we propose to compare their performances in identifying communities from Twitter data sets during natural disasters such as Hurricanes. Latent semantic analysis uses statistical computations, typically singular value decomposition, to find semantic or contextual meaning from the data. It finds relationships between terms and concepts in an unstructured data set. Probabilistic latent semantic analysis or indexing is another method based on Bayesian analysis that typically is used for two-mode data.;The objective is to compare these two methods on a large set of social media documents related to Hurricane Sandy in order to form clusters of similar concepts. We then compare the performance of these two methods to determine their relative performance in determining communities and hidden topics, e.g. finding clusters of similar topics like power outages, floods, gas outages, etc. We apply two clustering methods, K-Means and Affinity Propagation to form clusters in the data. Finally, we present the results by applying external methods of evaluation after creating a test data-set to compare the performance of these two methods. Metrics like Precision, Recall, confusion matrix are used to evaluate the performance of our system. The evaluation showed us that in almost all the scenarios, PLSA works better than LSA in finding out hidden relationships and structures. Whereas LSA is slightly faster than PLSA.