How to get started with machine learning techniques

I'm about a couple months into learning machine learning concepts and would really like general guidance on how to proceed with analysis' using R. My objective is to be able to look at some datasets in kaggle, and have a thought process of how I want to start my predictions. Not really interested in winning at this point, just want to learn.

Im reading a couple of machine learning books and there are many many techniques: decision trees, classification, neural networks, k-manifolds etc etc, and it is a bit overwhelming. Right now, I'm kind of focused on decision trees.

It seems like the folks at kaggle know which algorithm/techniques to apply and know how to troubleshoot to improve their results.

So is there a general outline/guidance on how to start with an analysis? I dont know which technique to use in which situation, and really don't know where to start.


Less is more. Stay pure. Stay poor.
Well a couple of months isn't that much time. Intuitive direction will come with practice. I also like the idea of machine learning. I think the first steps are just descriptive in nature and understanding what types of data you have. I would just start with the simpler methods and build up your repertoire.

The winners always seem to be Boosting or Neural Networks. I am no expert, but the traditional classification problems fit well with boosting models and the latent or identification problems seem better suited to NN. I have yet to really figure the later out. Just keep reading and watching videos. There seems to be a bottomless pit of resources out there on ML, so inundate yourself.
I'm about at the same stage than you are, interested in ML but its hard to get started, especially with no good background in mathematics.

I found the coursera Stanford course, its just started and i think you can still get started its only the second week. Unfortunately they are doing the assignments in Matlab or Octave both new to me but I'll try.
They guy in hlsmiths link mentiones that it might be too difficult but I think understanding gets better when writing the algorithms by yourself rather than only applying them (at least for me this was always the case).

If you have any valuable tips for me I would appreciate if you would share it =)


Probably A Mammal
I think it helps to have expectations about what you think ML consists of. For some people in the field, it's deep computer science algorithm development or code engineering. Others, it may just be to understand the difference techniques and how to apply them in their favorite program (R, SAS Enterprise Miner, Rapid Miner, etc.). They all have a common core of understanding the mathematical concepts, but not necessarily the exact math. You can do applied statistics, for instance, without really knowing statistical theory from a mathematical perspective ... but it helps! The more theory you know, the more you'll understand why things work, and why they don't work in this or that instance, and when faced with options, which choice is more appropriate.