What are the basic things to remember while using supervised learning?
Doing supervised learning is an art and involves more than just picking up a standard learning library and then dumping the whole data to it. There can be multiple approaches to achieve the desired accuracy. A few things which will surely help you once you are settled with the modeling technique are
1. Bias-Variance trade off
One needs to make a decision on how generic a model ought to be so that it has low variance when applied to various data sets. There is a need for both, low bias and low variance in the field of machine learning and this decision-choice depends on the problem one is solving.
2. The curse of dimensionality
Though it is good to have a lot of features to predict, the more the dimensions the difficult it gets for the model to be learnt.
3. Choosing smart features
This is one very important criterion which helps in making very stable models. Always go for features which you are sure won’t be noisy. Features which have a strong theoretical backing for their selection would generally be stable and should be preferred.
How does one deal with categorical data?
There are several ways of handling categorical data. If the categorical data is ordered, they could be used as other continuous variables according to their numerical values. If they are without any order, we can’t directly allocate them numerical values. For instance, let us say each data point has a color- green, yellow and blue. We cannot allocate them numbers such as 1, 2, 3, since that would imply green < yellow < blue, which is not the case. In such case the categorical variable is split into multiple variables: variable1: 0-1 (green or not green); variable2: 0-1 (yellow or not yellow). You do not need a third variable, since that happens when the first two variables are 0. Now you can use these two variables like ordered categorical data. Remember, this is just one way.
Why does one need to do factor analysis?
Factor analysis is a method for investigating whether a number of variables of interest Y1, Y2, :: :, Yl, are linearly related to a smaller number of un-observable factors F1, F2, : ::, Fk . It is used actively for dimensionality reduction of a data set- to reduce the number of variables, yet preserving the information in them. That may help make more generic models
What are Conservatives and Liberals
A Conservative predictor is a predictor which has a tighter filter on the number of poor performers accepted in the final pool, even if it means a few good performers are sacrificed in the process of eliminating the poor performers. While in the case of liberals, they want to lose as little good performers as possible. They are fine if a few poor performers are allowed in the final pool.
These two cases just demonstrate the most common situations we face in the industry in our projects. In case the supply of good candidates is high, companies prefer to have a conservative mindset, whereas in industries where the number of good candidates is low, companies are not willing to lose on any good performer.
How do you proceed when data is imbalanced?
Manage it in the cost function and do not use a simple use misclassification error.