{"id":139,"date":"2016-08-02T17:42:52","date_gmt":"2016-08-02T16:42:52","guid":{"rendered":"http:\/\/www.mub.eps.manchester.ac.uk\/in-abstract\/?p=139"},"modified":"2017-04-11T15:20:59","modified_gmt":"2017-04-11T14:20:59","slug":"cost-sensitive-boosting-algorithms-do-we-really-need-them","status":"publish","type":"post","link":"https:\/\/www.mub.eps.manchester.ac.uk\/in-abstract\/cost-sensitive-boosting-algorithms-do-we-really-need-them\/","title":{"rendered":"Cost Sensitive boosting Algorithms: Do we really need them?"},"content":{"rendered":"

A unifying framework to understand two decades of research.<\/strong><\/p>\n

The world relies on data, and increasingly on data analytics.\u00a0 \u201cBoosting\u201d is the name of a family of algorithms used extensively in data science, across numerous applications.\u00a0 \u201cCost sensitive\u201d algorithms are those that continue to work well when the cost of making a false positive prediction outweighs the cost of a false negative, or vice versa.\u00a0 This occurs, for example, in safety critical applications like medical informatics.<\/p>\n

It has been perceived that Boosting tends to fail (i.e. predict incorrectly) on such \u201cimbalanced cost\u201d data, hence much effort has been expended attempting to patch this perceived weakness.\u00a0\u00a0 This research identified 15 distinct Boosting variants, published over 20 years, each with its own claims to distinction and superiority.<\/p>\n

Researchers at Manchester analysed these, with 4 theoretical frameworks \u2013 Bayesian Decision Theory, Margin Theory, Functional Gradient Descent, and Probabilistic modelling – and axioms were identified, that algorithms must obey to be coherent.\u00a0\u00a0 The result was that only one<\/strong> algorithm is consistent with all frameworks, and practical for deployment in real applications.\u00a0 Surprisingly, this is the original 1997 algorithm, with a \u201ccalibration\u201d of its outputs.\u00a0 Extensive experiments support this, showing it outperforms all other variants.<\/p>\n

This research, in effect, renders irrelevant almost 20 years of heuristic algorithms, pointing out that a simple calibration of the original work is more than sufficient.<\/p>\n

\n