Mr. F
Reputation
1,524
Top tag
Next privilege 2,000 Rep.
Edit questions and answers
 Mar 28 comment Why does using a non-parametric test decrease power? @zero Another reason a person may choose a parametric method is for decision rule parsimony on future data. A simple linear regression with M covariates and N training samples only requires me to store between O(M) and O(M^2) data (the coefficients, intercept, and possibly covariances for standard errors), while most non-parametric models (for example, some kind of locally weighted regression) require me to keep O(N) data (the actual training samples themselves). For cases when N is extremely large, this can be impractical even if the non-parametric method is desirable. Mar 19 awarded Popular Question Feb 2 awarded Yearling Jan 25 awarded Popular Question Sep 22 awarded Guru Aug 17 comment How to assess the similarity of two histograms? This may not be as useful in image processing as in statistical fit assessment. Often in image processing, a histogram of data is used as a descriptor for a region of an image, and the goal is for a distance between histograms to reflect the distance between image patches. Little, or possibly nothing at all, may be known about the general population statistics of the underlying image data used to get the histogram. For example, the underlying population statistics when using histograms of oriented gradients would differ considerably based on the actual content of the images. Aug 9 comment How to compute the standard errors of a logistic regression's coefficients @jseabold However, if you want to get some ad hoc notion of feature importance in logistic regression, you cannot just read off the effect sizes (the coefficients) without thinking about their standard errors. So even if you're not doing a frequentist test, and you just want some indication of effect sizes and robustness, the sklearn lack of variance output is challenging. Jun 9 awarded Enlightened Jun 9 comment Can a -2 Log likelihood be calculated with only one model? @SibbsGambling In this link there is an example with coolibah tree data showing a "full" or "saturated" model where the log-likelihood is not zero. I believe there are certain situations where the saturated model must have a likelihood of one by definition, but not in all situations. Jun 9 awarded Nice Answer Apr 29 comment Having a job in data-mining without a PhD This is of course a generalization. Many Ph.D. level machine learning scientists are also quite gifted programmers and are happy to think in business-focused pragmatic lines. But on average, the extra time spent not interfacing algorithms with real world software-based needs does not necessarily confer any advantage over the equally-as-theoretically-capable masters-level graduates, and can certainly confer a disadvantage for folks whose over emphasis of theory has left them without a good, pragmatic software development skill set. Apr 29 comment Having a job in data-mining without a PhD In addition, this sort of expertise is not generally applicable in many applied settings, since it's often the case that organizations and corporations strongly prefer to use pre-existing algorithms and open-source implementations to enable faster prototyping and to avoid sinking time into in-house research that is too speculative to be a good use of money or time. I find that masters students, with a more pragmatic focus on gaining sufficient, but not excess, theoretical appreciation, tend to do better with this and are a bit more well-rounded in the software development that comes with it. Apr 29 comment Having a job in data-mining without a PhD @JackTwain I disagree. I know many masters-level machine learning scientists who stay very current on the professional literature, attend CVPR, NIPS, ICML, and other annual conferences, and have every bit the same (or even better) theoretical depth as any of the Ph.D. machine learning graduates that I know, whether talking about modern criticisms of class label noise for boosting classifiers, or PAC-learning work on the complexity of learning, many masters students spend a lot of time reading these things, both during school and after graduating. Feb 2 awarded Yearling Jan 31 comment Choosing right set of variables for Logistic regression and decision tree Have you considered choosing the variables x and y? I've seen a lot of people reporting good results for their regressions when they've chosen x and y. Sep 24 awarded Autobiographer Jul 9 comment Aside from regression coefficients, what are commonly used approaches to measure one variable's “sensitivity” to another variable? Last thought: I did show them the Achen paper that I linked above, which provides pretty comprehensive examples of how the boiler-plate regression approach can go wrong. They simultaneously acknowledged that it could go wrong and that there was not much theoretical reason why it should even work correctly in the first place ... and at the same time essentially said they did not care because they wanted something expressed as regression coefficients regardless of the ramifications of that sort of model. And these were highly educated veterans running a long-standing, successful company. :/ Jul 9 comment Aside from regression coefficients, what are commonly used approaches to measure one variable's “sensitivity” to another variable? I'm asking here to get feedback on the general problem. Such as, in machine learning sensitivity is measured via such-and-such, but in a classical Frequentist setting it is measured with blah-blah and in a Bayesian setting it's measured via blah... hopefully with references. Also, this is issue is quite old now and I have moved on from the position where the original project happened. I'm still very interested in this issue though ... just not connected to a specific problem instance anymore. Jul 2 awarded Curious Jun 16 revised Computing directly comparable wavelet features on variable-length training examples deleted 4 characters in body