2
$\begingroup$

I’ve always used highly flexible machine learning algorithms like boosted trees, support vector machines, and random forests that supposedly excel at identifying non-linear and irregular patterns and variable interactions. I’ve been building predictive algorithms for years with reasonable success, but I’ve never really invested much time in visualizing my data -- no scatter plots, charts, or bar graphs.

When working on a new problem, I try to understand my data by reading about the topic for many hours and absorbing other people’s subject matter expertise. People have analyzed my data for decades, so there is no shortage of expert opinions. When coming up with new ideas for predictors, am I better off reading more expert opinions because machine learning algorithms will generally "outperform my eyes"?

If I'm going to test the inclusion of each candidate variable based on the performance of out-of-sample data anyway, do visualization techniques like box plots become redundant? If a box plot says that a certain variable is important, wouldn't a boosted trees model or support vector machine reach the same conclusion?

My models are generally boosted trees. Are there certain visualization techniques that would help me see things that boosted trees could not?

$\endgroup$
  • $\begingroup$ See this page deltarho for effective visualization of complex data! $\endgroup$ – kjetil b halvorsen Jan 20 at 11:10
3
$\begingroup$

Firstly, could you clarify what you mean by 'highly flexible'?

Also, YES YES YES. Always do exploratory data analysis. Feature engineering is what will take your algorithms to the next level. Good EDA/plots will help you figure out good predictors to add.

$\endgroup$
  • $\begingroup$ By flexible I mean that the models don't require normally distributed data, and that they can uncover variable interactions without the user having to identify the variable interactions explicitly. Boosted trees are particularly good at identifying variable interactions $\endgroup$ – Ryan Zotti Feb 6 '15 at 2:59
  • $\begingroup$ Also, if I'm going to test the inclusion of each candidate variable based on the performance of out-of-sample data anyway, do visualization techniques like box plots become redundant? If a box plot says that a certain variable is important, wouldn't a boosted trees model reach the same conclusion? $\endgroup$ – Ryan Zotti Feb 6 '15 at 3:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.