Possible Duplicate:
Machine learning cookbook / reference card / cheatsheet?

There are numerous machine-learning approaches out there. Also there are numerous ways how to optimize their parameters and numerous ways to preprocess your data beforehand. It seems to be a common practice to initially implement some off-the-shelf approach (SVM, random forest) and then trying to beat the accuracy with more sophisticated approaches. I would like to know if there are any guidelines, recommendations or rules of thumb which approach to use in which context? Are recommendations data-specific or domain-specific (e.g. genetics, econometrics)? Also I would like to know how to account this multitude when writing results up. Trying out multiple approaches and only reporting the best one in your publication cannot be valid, right?

  • 1
    $\begingroup$ This question is perhaps too broad. Guessing what algorithms might work best comes down to understanding the nature of a data source and the workings of different algorithms to know which might work for the peculiarities of the data. The details of both parts (understanding the data and the algorithms) is literally the subject of many textbooks. I would recommend for instance 'The Elements of Statistical Learning', the entire contents of which would be a suitable answer for your question. Hopefully someone can provide you with a more succinct answer! $\endgroup$ – Bogdanovist Sep 5 '12 at 5:38
  • $\begingroup$ One more comment, it seems that for pure performance (as opposed to efficiency of calculation) the best approach is not to select the one 'best' method, but to create ensemble solutions employing a vast array of different techniques, possibly including everything that you have tried. This may not be appropriate in an academic study though, depending on what you are trying to achieve. $\endgroup$ – Bogdanovist Sep 5 '12 at 5:41