[英]SVM - Are there properties of the data that can indicate best parameters (e.g. C, gamma)
It seems pretty standard to use cross validation to determine the best parameters. 使用交叉验证来确定最佳参数似乎很标准。 Of course, this is usually a time-consuming process. 当然,这通常是一个耗时的过程。 Are there any shortcuts? 有捷径吗? Are there other, faster, forms of exploratory analysis that can provide a hint as to which values will be best? 还有其他更快的探索性分析形式可以提示哪些值是最佳的吗?
For example, at my current understanding of machine learning and SVM, I might do something like perform an initial grid search in the range of [10e-5, 10e5] at exponents of 10 for C, and then fine tune from there. 例如,以我目前对机器学习和SVM的理解,我可能会做一些类似的事情,以C的指数10在[10e-5,10e5]范围内执行初始网格搜索,然后从那里进行微调。 But is there a way I could quickly estimate that the best C is somewhere between 10e3 and 10e5, and then perform more specific searches? 但是有没有一种方法可以快速估算出最佳C在10e3和10e5之间,然后执行更具体的搜索?
This question probably applies to most ML techniques, but I happen to be working with SVM right now. 这个问题可能适用于大多数ML技术,但是我恰好正在使用SVM。
Yes, this is an area of active research! 是的,这是一个活跃的研究领域! There has been a lot of work in different approaches to hyper-parameter tuning besides the standard grid search we all know and (maybe?) love. 除了众所周知的(也许是?)热爱的标准网格搜索之外,还有许多方法可以用于超参数调整。
The area most similar to what you are describing are various bayesian / gaussian process approaches to the problem. 与您所描述的最相似的领域是解决问题的各种贝叶斯/高斯过程方法。 This github repo has an implementation and some informative pictures on how it works https://github.com/fmfn/BayesianOptimization . 这个github回购有一个实现和一些有用的图片,说明它是如何工作的https://github.com/fmfn/BayesianOptimization 。 This approach works by treating the parameter optimization problem as another machine learning problem, where we have features for every hyperparameter, and try to predict the performance of various parameter combinations. 这种方法通过将参数优化问题视为另一个机器学习问题来工作,我们在其中具有每个超参数的特征,并尝试预测各种参数组合的性能。
That is a high level description of the process, you can read the linked papers/notebooks in the repo for more details. 那是对该过程的高级描述,您可以阅读回购中的链接论文/笔记本以获取更多详细信息。
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.