How to make classification results replicable?

Question

Let's suppose I have a simple classification problem:

data={{1,"a"}->True,{2,"b"}->True,{3,"b"}->True,{4,"a"}->True, {5,"a"}->False,{6,"b"}->True};
c=Classify[data, Method->"RandomForest"];
c[{5, "b"},"Probabilities"]

This piece of code produces different results occasionally:

(*<|False->0.294872,True->0.705128|>
<|False->0.335242,True->0.664758|>
<|False->0.271804,True->0.728196|>*)

I guess it happens because of internal pseudorandom generator calls within RandomForest. How can I make my results replicable (possible give a seed to internal generator)?

Setting a seed using SeedRandom does not make the result deterministic.

Answer 1

You can save your trained classifier in .mx file and simply load it in the new session.

data = {{1, "a"} -> True, {2, "b"} -> True, {3, "b"} -> 
    True, {4, "a"} -> True, {5, "a"} -> False, {6, "b"} -> True};
c = Classify[data, Method -> "RandomForest"];
c[{5, "b"}, "Probabilities"]

<|False -> 0.202281, True -> 0.797719|>

Export["c.mx", c];

Quit[]

Classify; 

c = Import["c.mx"];

c[{5, "b"}, "Probabilities"]

<|False -> 0.202281, True -> 0.797719|>

Notice that you need to call Classify; before importing.