The Comparison of Effects of Relevant-Feature Selection Algorithms on Certain Social-Network Text-Mining Viewpoints

摘要

This research addresses a well-known problem in the area of text mining: The high computational complexity caused by many irrelevant features (terms, words), which may play an appreciable role of noise from the classification point of view and non-linearly rule the time and memory requirements. Using a set of real-world textual documents represented by sentiment related to three selected and extensively tracked Internet sources freely written in English, a group of available algorithms (Gain Ratio, Chi Square, Info Gain, Symmetrical Uncertainty, Winnow, One R, Relief F, Principal Components, SVM, LSA) applied to discovering relevant features was tested with 10,000, 25,000, and 50,000 social-network entries. All the algorithms provided very similar results concerning looking for the relevant features - typically, only the feature significance rank was slightly different. Except for some slower algorithms, the term-preselecting time ranged from seconds to minutes to a couple of hours. However, after using only a relevant fraction of features instead of all of them, the entry length very considerably decreased by several orders of magnitude, particularly for larger data sets having very high dimensionality degree. Despite the extremely strong reduction of the number of words, the classification accuracy remained the same independently on the relevant-feature selection algorithm choice.