Analyzing the presence of noise in multi-class problems: alleviating its influence with the One-vs-One decomposition

摘要: The presence of noise in data is a common problem that produces several negative consequences classification problems. In multi-class problems, these are aggravated terms accuracy, building time, and complexity the classifiers. cases, an interesting approach to reduce effect decompose into binary subproblems, reducing and, consequently, dividing effects caused by each subproblems. This paper analyzes usage decomposition strategies, more specifically One-vs-One scheme, deal with noisy datasets. order investigate whether able or not, large number datasets created introducing different levels types noise, as suggested literature. Several well-known algorithms, without decomposition, trained on them check when advantageous. results obtained show methods using strategy lead better performances robust classifiers dealing data, especially most disruptive schemes.