Why does the decision tree return different solutions for the exact same training data

Question

I was trying out a ML example and it worked for the most part but when I ran the code consecutively python started spitting out different prediction results, now I am now ML expert but that seems wack?

# Example file from Google Developers: "Hello World - Machine Learning Recipes": YouTube: https://youtu.be/cKxRvEZd3Mw
# Category: Supervised Learning                                                                               
# January 14, 2018                                                                                            
from sklearn import tree                                                                                      

# Declarations: Texture                                                                                        
bumpy = 0                                                                                                      
smooth = 1                                                                                                     

# Declarations: Labels                                                                                         
apple = 0                                                                                                      
orange = 1                                                                                                                                                                 

# Step(1): Collect training data                                                                               
# Features: [Weight, Texture]                                                                                  
features = [[140, smooth], [130, smooth], [150, bumpy], [170, bumpy]]                                          

# labels will be used as the index for the features                                                            
labels = [apple, apple, orange, orange]                                                                        

# Step(2): Train Classifier: Decision Tree                                                                     
# Use the decision tree object and then fit 'find' paterns in features and labels                              
clf = tree.DecisionTreeClassifier()                                                                            
clf = clf.fit(features, labels)                                                                                

# Step(3): Make Predictions                                                                                    
# the prdict method will return the best fit from the decesion tree                                            
result = clf.predict([[150, bumpy], [130, smooth], [125.5, bumpy], [110, smooth]])                             
# result = clf.predict([[150, bumpy]])                                                                         
print("Step(3): Make Predictions: ")                                                                           
for x in result:                                                                                               
    if x == 0:
    print("Apple")                                                                                        
        continue                                                                                              
    elif x == 1:                                                                                              
        print("Orange")                                                                                       
        continue                                                                                              
    print("Orange")                                                                                        

Click link to see vim and bash windows

Answer 1

There's an element of randomness to (most?) Decision Tree algorithms, and your training set is very small which might be exaggerating the effect. The randomness is typically used to determine how many/which samples to use, and in your case there are very few samples.

Try setting the random_state to some fixed integer when you create the DecisionTreeClassifier. If you want a repeatable result for testing, you'll need to use the same "seed" value each time. They use a random seed of zero in the example docs:

clf = DecisionTreeClassifier(random_state=0)