I have not found any clear explanation about the necessity of checking correlation between features when doing classification.
Is it necessary to check the correlation when the accuracy of classification is satisfactory?
Zeynab Mousavikhamene without feature selection and feature extraction or attribute selection, you can perform classification but, results will not be accurate. if you select featurre or attribute, results will be focused or concentrated.
Feature/attribute correlation is an important step in the feature selection/reduction phase of the data pre-processing especially for the data type for which the features are continuous.
Hi!!
This discussion may help you : https://towardsdatascience.com/feature-selection-correlation-and-p-value-da8921bfb3cf?gi=26e9b0d7db7a
If the idea is to reduce redundancy due to so many features or a problem of large datasets reducing the speed of classification, then you must continually check the correlation till your are satisfied. there must be a balance between classification and the amount of feature used
It is not necessary to check the correlation between features in order to create a classification model, but if you know how to do that, you should. If you have many features and most of them are irrelevant for your predictions, your model will probably not be able to reach its full potential.
By understanding which data are related to what you are trying to predict can significantly help your model achieve higher accuracy. At the same time training time will be reduced because of the reduced dimensions of data.
Let me give you a simple example. Let's say you want to predict whether a client of a car insurance company will cost money based on his/her characteristics. You already have information about past clients and whether they requested money to cover their car expenses or not. The data you have for the clients are: name, surname, age, years of driving experience, number of past accidents, marital status, address, etc. Some of these information will be very useful for your model, like age and years of experience, while others will not help your model at all, like name and surname. So you will be better leaving those information our of your training data.
If you are not sure if a feature is relative or not then it would be better if you include it, because you do not want to loose helpful information. If you can use all the features for training your classification model give it a try, then remove the ones you believe that are irrelevant and try again. You can compare the two results and decide which model to use, based on their accuracy.
We must not forget that our models can sometimes find a correlation between some features that we could not spot with data analysis. That is the reason that we are using these models after all, isn't it?
I have done feature-analysis in some of my articles after (sic!) I got a sufficient result with the classification, mainly as an insight into the features, to explain the reader and myself which features were most important. As I see, you also have done the classification already. So, you, first, hypothesized about which features to use, then used them, and only after that you came up with the question: how do I check which features were correct to use? I think, presently we mainly rely on the expert opinion about the features to build classifiers, because the statistical instruments are not ready to use for the modern classification tasks, as data science and computer methods of data aggregation are developing very fast. I tend to think there are either yet no methods to bind correlation of features to classification performance or these methods are poorly rationalized and, hence, not reliable. Therefore, if you choose to do statistical analysis of features, you will end up with the rationalization of the statistical method first. However, the effort might pay off and you will contribute to the field. We try to do something like that in the pre-print I attach below.
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification