Learning some Machine Learning (while learning from it...)

Machine learning is ingrained in our day-to-day life. It is part of our spam filters mechanism, voice command smartphone interpretation and any search on Google. Chances are good that machine learning has been helping you along somewhere in you life.

But what is machine learning? 

Machine learning exists at the intersection of computer science and statistics.

That might be true... but a little too deep to begin with. Let's start with some basic machine learning concepts :

Problem Definition

Before looking at machine learning models or even start with data collection, one should define well the problem needs to be solved. Remember that eventually, a computer program should look into data , measure values and predict some results.  A clear problem definition will prevent using the wrong machine learning tools or data set.

Clean / Transform Data

Preparing and understanding the data set before using it is always important. It becomes critical  when dealing with machine learning and big data. Each small mistake can lead to a huge impact on the expected results. More data is not always better results.

There are times when more data helps; there are times when it doesn't.

Since the sample size effects the computation resource requirement, there are times when more data helps; there are times when it doesn't.

Analyzing, preparing and cleaning the data is a big portion of what eventually results in good machine learning... 

Pick up the right tool 

Machine learning is a group of tools and techniques for multiple type of problems. Picking up the right tool is a essential, after the problem and existing data set is well defined. Remember, machine learning is always about estimation and making the best guess, so don't expect perfect results. There is always margin of error, noise,  correlation coefficient and others... Machine learning is also about trial and error. 

Here are few technical terms you should be aware of: 

Supervised learning - When data points have "labels" assigned to them to "teach" the expected output per input. The algorithm will train on the labeled data and predict labels for new inputs.  

Unsupervised learning  - no labels are available for training  the algorithm, leaving it on its own to find structure in its input.

Classification vs Regression 

Classification is when the results should be 1 of n values, and each wrong prediction is equally wrong. For example, if you're trying to classify images of items, identify a cat as a house isn't any better or worse than identify a dog as a house. With classification the output variable takes class labels (in our example - house, cat, dog...).

Machine learning focuses on prediction, based on known properties learned from the training data.

Regression is used when there's some sense of distance between the values. For example, if the actual value of market stock is 150$ and you predicted it to be $149.4, that's a pretty good prediction, while $10 is a much worse prediction. With regression the output variable takes continuous values (in our example what should be the market stock value).

Clustering

Similar to classification in clustering we're trying to group the data, only that the data is not labelled before hand. Clustering look at correlation and see if the data can be divided into groups based on similarity. 

Since clustering has no predefined labels, it uses unsupervised learning methods for the training period.

Dimensionality Reduction

Many modern machine learning problems take multiple dimensions of data to build predictions using many coefficients. Dimensionality Reduction simplifies data processing by mapping them into a lower-dimensional space. In many cases non-numeric values should be converted to number values before going on a dimensional reduction phase. 

We should stop here... (it's getting out of control already :-)). Many smart people keep on developing machine learning algorithm and spend their entire lifetime on just studying what's already there... 

Present the results / Predict 

Validate

Nothing much to say about that. Whatever results you get, validate those on real "outside" data. Internal testing can't replace an actual production environment.

Simplify

In most cases machine learning "consumers" are not data scientist or have any expertise in statistics or even knowledge of the data set. The only think they want is an answer. Example of results could be "77 with 30% margin error", "10/90 ratio", "True / False" and so on. Try playing with the results and present them in simple English words vs convoluted formulas or meaningless numbers.   

Improve Results

Consider where the model does not work well or what parts the model does not answer.  Go back to the initial problem definition and compare it with the results. Most machine learning algorithms, can accept reinforcement and adjustments parameters, for improving the results for future predictions. 

Reporting

Like most fortune-tellers know, presenting the prediction is as important as the prediction itself. Visualization is one of the best ways to present machine learning results. Interactive reports with dashboard and drill down capabilities, allow a better understanding of the results.

 Before teaching us anything, machine learning should "learn". As such the problem definition, data cleanup, model usage and presentation should be well implemented. The results could be not less than amazing.