which one is the best traditional machine learning algorithm? which one is more revolutionary?
w*g
2 楼
in terms of practical problem solving, I would say SVM is better. In theory, the idea behind boosting, i.e. weak learner vs strong learner, is a fundamental advancement in machine learning theory, probably of philosophical importance. SVM itself is more technical, but the theory that remotely backs SVM, i.e. VC theory, is even more fundamental and more important than boosting. However, the trivial version of SVM, i.e. linear SVM, has gained so much attention lately that people view SVM more often simply as "large margin" regression, rather than dimension reduction with support vectors as is the original intention of SVM. Both algorithms are flawed in a similar way in practice: both are meta-algorithms that rely on a user-defined plug-in: Boosting requires a decision stump, and SVM requires a kernel. Neither theory provides any insight on how the plugin must be designed. Both are revolutionary in its own way. But if we limit to the algorithms and do not generalize to as much as VC theory, I would say boosting is more revolutionary. If I have to pick one from the two to solve a problem, I would pick SVM for its performance. 着两个算法背后的大神, Robert Schapire和Vladimir Vapnik都在Princeton. 后者在NEC lab, 据说有一阵子年年申请Princeton大学的职位年年被毙, 大概是 因为学校觉得他不可能做出更大的成就了. 据他自己说是因为他是俄国人被歧视 了. Robert Schapire每年都在系里卖她女儿的童子军饼干. 这哥们开一门理论 机器学习的课, 有一年我还被拉去做他的助教. 这课就是推一个学期的公式, 最后some how证明SVM和boosting在他的框架下其实是一回事. 在 nostring (尼) 的大作中提到: 】