如何从决策树中对功能进行取消编码以查看重要功能？

Question

I have a dataset that I am working with. I am converting them from categorical features to numerical features for my decision tree. The conversion happens on the entire data frame with the following lines:

le = LE()
df = df.apply(le.fit_transform)

I later take this data and split it into training and testing data with the following:

target = ['label']
df_y = df['label']
df_x = df.drop(target, axis=1)

# Split into training and testing data
train_x, test_x, train_y, test_y = tts(df_x, df_y, test_size=0.3, random_state=42)  

Then I am passing it to a method to train a decision tree:

def Decision_Tree_Classifier(train_x, train_y, test_x, test_y, le):
    print " - Candidate: Decision Tree Classifier"
    dec_tree_classifier = DecisionTreeClassifier(random_state=0) # Load Module
    dec_tree_classifier.fit(train_x, train_y) # Fit
    accuracy = dec_tree_classifier.score(test_x, test_y) # Acc
    predicted = dec_tree_classifier.predict(test_x)
    mse = mean_squared_error(test_y, predicted)

    tree_feat = list(le.inverse_transform(dec_tree_classifier.tree_.feature))
    print "Tree Features:"
    print tree_feat
    print "Tree Thresholds:"
    print dec_tree_classifier.tree_.threshold

    scores = cross_val_score(dec_tree_classifier, test_x, test_y.values.ravel(), cv=10)
    return (accuracy, mse, scores.mean(), scores.std())

In the above method, I am passing the LabelEncoder object originally used to encode the dataframe. I have the line

tree_feat = list(le.inverse_transform(dec_tree_classifier.tree_.feature))

To try and convert the features back to their original categoric representation, but I keep getting this stack trace error:

  File "<ipython-input-6-c2005f8661bc>", line 1, in <module>
    runfile('main.py', wdir='/Users/mydir)

  File "/Users/me/anaconda2/lib/python2.7/site-packages/spyder_kernels/customize/spydercustomize.py", line 668, in runfile
    execfile(filename, namespace)

  File "/Users/me/anaconda2/lib/python2.7/site-packages/spyder_kernels/customize/spydercustomize.py", line 100, in execfile
    builtins.execfile(filename, *where)

  File "/Users/me/mydir/main.py", line 125, in <module>
    main()  # Run main routine

  File "candidates.py", line 175, in get_baseline
    dec_tre_acc = Decision_Tree_Classifier(train_x, train_y, test_x, test_y, le)

  File "candidates.py", line 40, in Decision_Tree_Classifier
    tree_feat = list(le.inverse_transform(dec_tree_classifier.tree_.feature))

  File "/Users/me/anaconda2/lib/python2.7/site-packages/sklearn/preprocessing/label.py", line 281, in inverse_transform
    "y contains previously unseen labels: %s" % str(diff))

ValueError: y contains previously unseen labels: [-2]

What do I need to change to be able to look at the actual features themselves?

Answer 1

When you do this:

df = df.apply(le.fit_transform)

you are using a single LabelEncoder instance for all of your columns. When called fit() or fit_transform() , le will forget the previous data and learn the current data only. So the le you have is only storing the information about the last column it seen, not all columns.

There are multiple ways to solve this:

1. You can maintain multiple LabelEncoder objects (one for each column). See this excellent answer here:

   • Label encoding across multiple columns in scikit-learn

      from collections import defaultdict d = defaultdict(LabelEncoder) df = df.apply(lambda x: d[x.name].fit_transform(x)) 

2. If you want to keep a single object to handle all columns, you can use the OrdinalEncoder if you have the latest version of scikit-learn installed.

    from sklearn.preprocessing import OrdinalEncoder enc = OrdinalEncoder() df = enc.fit_transform(df) 

But still the error will not be solved, because the tree_.feature dont correspond to values of the features, but the index (column in df ) that was used for splitting at that node. So if you have 3 features (Columns) in the data (irrespective of values in that column), the tree_.feature can have values:

• 0, 1, 2, -2

• -2 is a special placeholder value to denote that the node is a leaf node, and so no feature is used to split anything.

tree_.threshold will contain the values corresponding to your values of data. But that will be in floats, so there you will have to convert according the conversion of categories to numbers.

See this example for understanding the tree structure in detail:

• https://scikit-learn.org/stable/auto_examples/tree/plot_unveil_tree_structure.html