# Applying machine learning in a real world example

I've almost finished the Stanford Machine Learning Course from Andrew Ng. Now I think I am capable of applying algorithms like logistic regression, linear regression, Neural Networks and so on. To practice, I started to look at Titanic Kaggle dataset. I was thinking to apply a Logistic Regression method to the problem.

As far as I know, I think I need to make a descriptive analysis of my dataset, for example, to choose the features set. The Andrew Ng course didn't cover this.

Whats the importance of make a Descriptive Analysis of dataset before start to train an ML algorithm?

Or

Is it best to apply a "simple and dirty" ML algorithm to the dataset and start to refine?

I was thinking in this course.

• Btw, If you are looking for more online courses, consider this one r-bloggers.com/… and as a comment: "ML" is ambiguous, in stats it is used for both "machine learning" and "maximum likelihood", so better use the full name (I edited your title). – Tim Sep 25 '15 at 13:32
• That's a wonderful link @Tim. – Dawny33 Sep 25 '15 at 13:33

Whats the importance of doing Descriptive Analysis of dataset before start to train an ML algorithm?

You might want to refer to this question which discusses the difference between Descriptive Statistics and Exploratory statistics (or analysis).

Descriptive statistics helps in extracting as much useful information from the dataset as possible. So, it is done before the model is constructed.

Is it best to apply a "simple and dirty" ML algorithm to the data set and start to refine?

No, just throwing any algorithm at the dataset might give you some accuracy and some results, but you wouldn't learn anything by doing so.

Instead, start by taking a numerical dataset, apply linear regression on it, and learn about the importance of coefficients and their effect on the regression model.

Then, try to optimize the results with a basic Gradient Descent algorithm. Appreciate the beauty of it. The concept of moving down the curve with differentials, is really good for an absolute beginner to savour the concept of predictive analytics and ML.

Then, take up a classification problem. Kaggle's titanic problem is a nice start.

The book on Pattern Recognition and Machine Learning by Bishop is an excellent book to have.

Some really nice ML problems on real-world datasets for getting hands dirty:

• Kaggle's Bicycle Demand: One has to forecast bike rental demand depending on the features in the data.
• US Census data: Working on this dataset would help you get really good at doing descriptive statistics.
• Analytics Edge on EdX: A nice course on EdX about real world analytics, which ends with a Kaggle competition for the course participants.
• In the Andre Ng i already implemented (using matlab) a grandient descent and other algorithms like linear regression and logisc regression. Andrew Ng suggest always make a "simple and dirty" implementation and the refine. But IMO, i need to make D/E Analisys, (for example, features selection) and then start to apply "simple and dirty" ML. – p.magalhaes Sep 25 '15 at 12:59
• Yes, you have to start with the descriptive analytics part. And by "simple and dirty" doesn't mean throwing just any algorithm at the data. For example: You might not understand a k-means clustering algo. now. So, it doesn't make sense to use it on the data. So, use a simple and a starter level algo.; and then build up your principles from it. – Dawny33 Sep 25 '15 at 13:26

I posted the same question on Course Machine Learning forum discussion, and here is a response from one of those mentor:

"No analysis of the system is needed. It may be useful, that is up to you. You may with to apply many different ML techniques and see which ones give the best predictions.

One method of exploratory analysis is by observing the theta values after training the system. Theta values that are near zero have little impact on the system, for example."

Tom Mosher

• Yeah, the mentor is perfectly right. No analysis of the system is needed: Yes, exploratory and descriptive statistics are optional, but are highly recommended. theta values are the coefficients of the regression model, thus adding to my "importance of coefficients" line. – Dawny33 Sep 26 '15 at 14:49