There are many different algorithms for feature selection and also for classification tasks. I want to know what feature selection algorithm works better, for my case, to be used in a classification task. Should I decide first the feature selection algorithm using a unique classification method? Could I test many combinations of different feature selection and classification algorithms at a time and see which one give me the best accuracy on a test set?
I would appreciate any help,
Thank you!
Dear Priscila Marcondes ,
You should check this blog, he explains very details of feature selection, classification, how to use and choose the best algorithm for your problem.
https://machinelearningmastery.com/an-introduction-to-feature-selection/
This question is dependent to many factors. Having domain knowledge may help you to select appropriate features. If you don't have deep domain knowledge, after having a rich set of features data there is a rich set of feature selection algorithms including RFE (Recursive Feature Elimination), ReliefF, LDA ( Linear Discriminant Analysis), ANOVA ( Analysis of Variance),
mRMR (maximum relevancy-minimum redundancy), PCA ( Principal Component Analysis) etc. This question is also related to your problem. For example, here different methods have been studied for human voice features :
https://www.researchgate.net/publication/326280450_Comparison_of_feature_selection_methods_in_voice_based_emotion_recognition_systems In case of text classification tasks ,TF ( Term Frequency), DF ( Document Frequency), TF-DF etc. like approaches are commonly used :
https://www.researchgate.net/publication/326300968_Applying_Machine_Learning_and_Natural_Language_Processing_Techniques_to_Twitter_Sentiment_Classification_for_Turkish_and_English Especially in image processing and unstructured data problems unsupervised feature extraction methods are also commonly used. Extracting features with unsupervised module like CNN help you to extract also hidden relations.
In additions to the previous answers, as you said, there are many different algorithms for features, and there are many classifiers.. So, there is no algorithm/classifier can be used as the best. In order to do so, you have test tens of algorithms and with each one a single classifier should be used. We are talking about hundreds/thousands of combinations.
This is a hard or may be impossible job. Because the algorithm X with the Classifier C, needs for time t for checking and analysing. If we are talking about wrapper feature selection.
There is something else didn't mention, which is the model selection algorithm, or sometimes it is called "Automated Machine learning models". It is an algorithm searches for the best machine learning model/feature selection for a given dataset. One of the most popular frameworks is "TPot " in Python. You may read the following articles..
https://towardsdatascience.com/tpot-automated-machine-learning-in-python-4c063b3e5de9
https://www.datacamp.com/community/tutorials/tpot-machine-learning-python
https://github.com/EpistasisLab/tpot
Hope this helps,
Regards ..
Sinan Q. Salih , thank you for the answer and the great recommendation! I didn´t know about this Automated Machine Learn models! I was wondering if this kind of search for the best feature selection/classifier combination could give some false positive results and need a multiple testing correction, do you know it?
Best Regards,
Priscila
in the case of classification, there is not a valid algorithm for all the problems. So in my opinion you can use several algorithms, after making a comparison in terms of results and complexity.
- Badges
- Science topic