Which are the most important parameters to tune for k-NN, NB, RF, DL, SVM for text classification?

jochen_hartmann

Dear community,

 

I would like to compare the performance of the following five algorithms on different text classification tasks*:

 

1. k-Nearest Neighbors (k-NN)
2. Naive Bayes (NB)
3. Random Forest (RF)
4. Deep Learning (DL)
5. Support Vector Machines (SVM)

 

Question 1: Which paramesters are the most important to optimize for each method 1-5?

Question 2: What ranges should I give those parameters in the parameter optimization operator in order to avoid "boiling the ocean"?

 

Thanks in advance!

 

* each task has between 3 to 5 classes and the text length varies between 3 to 70 words per document / example

Thomas_Ott

Great question!

 

1. With K-nn I would optimize around "k".
2. Naive Bayes I usually don't optimize
3. Random Forest I would optimize depth of trees, # of trees, confidence
4. Deep Learning I'm not sure but I would choose a few of the activation functions
5. For text, I would use a LinearSVM and optimize C. 

Telcontar120

Excellent suggestions from @Thomas_Ott as usual.  I would add a couple more:

• There isn't actually anything to optimize with Naive Bayes, there is only one parameter (Laplace correction) and I would definitely leave it on.
• For Random Forest, I would also optimize the growing criterion (information gain, gain ratio, Gini, accuracy).
• For SVM, you might also try a polynomial kernel and optimize C as well as degree in the range of 1-4.