How to extract the decision rules from scikit-lear

问题:

Can I extract the underlying decision-rules (or 'decision paths') from a trained tree in a decision tree as a textual list?

Something like:

if A>0.4 then if B<0.2 then if C>0.8 then class='X'

Thanks for your help.

回答1:

I believe that this answer is more correct than the other answers here:

from sklearn.tree import _tree

def tree_to_code(tree, feature_names):
    tree_ = tree.tree_
    feature_name = [
        feature_names[i] if i != _tree.TREE_UNDEFINED else "undefined!"
        for i in tree_.feature
    ]
    print "def tree({}):".format(", ".join(feature_names))

    def recurse(node, depth):
        indent = "  " * depth
        if tree_.feature[node] != _tree.TREE_UNDEFINED:
            name = feature_name[node]
            threshold = tree_.threshold[node]
            print "{}if {} <= {}:".format(indent, name, threshold)
            recurse(tree_.children_left[node], depth + 1)
            print "{}else:  # if {} > {}".format(indent, name, threshold)
            recurse(tree_.children_right[node], depth + 1)
        else:
            print "{}return {}".format(indent, tree_.value[node])

    recurse(0, 1)

This prints out a valid Python function. Here's an example output for a tree that is trying to return its input, a number between 0 and 10.

def tree(f0):
  if f0 <= 6.0:
    if f0 <= 1.5:
      return [[ 0.]]
    else:  # if f0 > 1.5
      if f0 <= 4.5:
        if f0 <= 3.5:
          return [[ 3.]]
        else:  # if f0 > 3.5
          return [[ 4.]]
      else:  # if f0 > 4.5
        return [[ 5.]]
  else:  # if f0 > 6.0
    if f0 <= 8.5:
      if f0 <= 7.5:
        return [[ 7.]]
      else:  # if f0 > 7.5
        return [[ 8.]]
    else:  # if f0 > 8.5
      return [[ 9.]]

Here are some stumbling blocks that I see in other answers:

1. Using tree_.threshold == -2 to decide whether a node is a leaf isn't a good idea. What if it's a real decision node with a threshold of -2? Instead, you should look at tree.feature or tree.children_*.
2. The line features = [feature_names[i] for i in tree_.feature] crashes with my version of sklearn, because some values of tree.tree_.feature are -2 (specifically for leaf nodes).
3. There is no need to have multiple if statements in the recursive function, just one is fine.

回答2:

I created my own function to extract the rules from the decision trees created by sklearn:

import pandas as pd
import numpy as np
from sklearn.tree import DecisionTreeClassifier

# dummy data:
df = pd.DataFrame({'col1':[0,1,2,3],'col2':[3,4,5,6],'dv':[0,1,0,1]})

# create decision tree
dt = DecisionTreeClassifier(max_depth=5, min_samples_leaf=1)
dt.fit(df.ix[:,:2], df.dv)

This function first starts with the nodes (identified by -1 in the child arrays) and then recursively finds the parents. I call this a node's 'lineage'. Along the way, I grab the values I need to create if/then/else SAS logic:

def get_lineage(tree, feature_names):
     left      = tree.tree_.children_left
     right     = tree.tree_.children_right
     threshold = tree.tree_.threshold
     features  = [feature_names[i] for i in tree.tree_.feature]

     # get ids of child nodes
     idx = np.argwhere(left == -1)[:,0]     

     def recurse(left, right, child, lineage=None):          
          if lineage is None:
               lineage = [child]
          if child in left:
               parent = np.where(left == child)[0].item()
               split = 'l'
          else:
               parent = np.where(right == child)[0].item()
               split = 'r'

          lineage.append((parent, split, threshold[parent], features[parent]))

          if parent == 0:
               lineage.reverse()
               return lineage
          else:
               return recurse(left, right, parent, lineage)

     for child in idx:
          for node in recurse(left, right, child):
               print node

The sets of tuples below contain everything I need to create SAS if/then/else statements. I do not like using do blocks in SAS which is why I create logic describing a node's entire path. The single integer after the tuples is the ID of the terminal node in a path. All of the preceding tuples combine to create that node.

In [1]: get_lineage(dt, df.columns)
(0, 'l', 0.5, 'col1')
1
(0, 'r', 0.5, 'col1')
(2, 'l', 4.5, 'col2')
3
(0, 'r', 0.5, 'col1')
(2, 'r', 4.5, 'col2')
(4, 'l', 2.5, 'col1')
5
(0, 'r', 0.5, 'col1')
(2, 'r', 4.5, 'col2')
(4, 'r', 2.5, 'col1')
6

回答3:

I modified the code submitted by Zelazny7 to print some pseudocode:

def get_code(tree, feature_names):
        left      = tree.tree_.children_left
        right     = tree.tree_.children_right
        threshold = tree.tree_.threshold
        features  = [feature_names[i] for i in tree.tree_.feature]
        value = tree.tree_.value

        def recurse(left, right, threshold, features, node):
                if (threshold[node] != -2):
                        print "if ( " + features[node] + " <= " + str(threshold[node]) + " ) {"
                        if left[node] != -1:
                                recurse (left, right, threshold, features,left[node])
                        print "} else {"
                        if right[node] != -1:
                                recurse (left, right, threshold, features,right[node])
                        print "}"
                else:
                        print "return " + str(value[node])

        recurse(left, right, threshold, features, 0)

if you call get_code(dt, df.columns) on the same example you will obtain:

if ( col1 <= 0.5 ) {
return [[ 1.  0.]]
} else {
if ( col2 <= 4.5 ) {
return [[ 0.  1.]]
} else {
if ( col1 <= 2.5 ) {
return [[ 1.  0.]]
} else {
return [[ 0.  1.]]
}
}
}

回答4:

There is a new DecisionTreeClassifier method, decision_path, in the 0.18.0 release. The developers provide an extensive (well-documented) walkthrough.

The first section of code in the walkthrough that prints the tree structure seems to be OK. However, I modified the code in the second section to interrogate one sample. My changes denoted with # <--

Edit The changes marked by # <-- in the code below have since been updated in walkthrough link after the errors were pointed out in pull requests #8653 and #10951. It's much easier to follow along now.

sample_id = 0
node_index = node_indicator.indices[node_indicator.indptr[sample_id]:
                                    node_indicator.indptr[sample_id + 1]]

print('Rules used to predict sample %s: ' % sample_id)
for node_id in node_index:

    if leave_id[sample_id] == node_id:  # <-- changed != to ==
        #continue # <-- comment out
        print("leaf node {} reached, no decision here".format(leave_id[sample_id])) # <--

    else: # < -- added else to iterate through decision nodes
        if (X_test[sample_id, feature[node_id]] <= threshold[node_id]):
            threshold_sign = "<="
        else:
            threshold_sign = ">"

        print("decision id node %s : (X[%s, %s] (= %s) %s %s)"
              % (node_id,
                 sample_id,
                 feature[node_id],
                 X_test[sample_id, feature[node_id]], # <-- changed i to sample_id
                 threshold_sign,
                 threshold[node_id]))

Rules used to predict sample 0: 
decision id node 0 : (X[0, 3] (= 2.4) > 0.800000011921)
decision id node 2 : (X[0, 2] (= 5.1) > 4.94999980927)
leaf node 4 reached, no decision here

Change the sample_id to see the decision paths for other samples. I haven't asked the developers about these changes, just seemed more intuitive when working through the example.

回答5:

from StringIO import StringIO
out = StringIO()
out = tree.export_graphviz(clf, out_file=out)
print out.getvalue()

You can see a digraph Tree. Then, clf.tree_.feature and clf.tree_.value are array of nodes splitting feature and array of nodes values respectively. You can refer to more details from this github source.

回答6:

Scikit learn introduced a delicious new method called export_text in version 0.21 (May 2019) to extract the rules from a tree. Documentation here. It's no longer necessary to create a custom function.

Once you've fit your model, you just need two lines of code. First, import export_text:

from sklearn.tree.export import export_text

Second, create an object that will contain your rules. To make the rules look more readable, use the feature_names argument and pass a list of your feature names. For example, if your model is called model and your features are named in a dataframe called X_train, you could create an object called tree_rules:

tree_rules = export_text(model, feature_names=list(X_train))

Then just print or save tree_rules. Your output will look like this:

|--- Age <= 0.63
|   |--- EstimatedSalary <= 0.61
|   |   |--- Age <= -0.16
|   |   |   |--- class: 0
|   |   |--- Age >  -0.16
|   |   |   |--- EstimatedSalary <= -0.06
|   |   |   |   |--- class: 0
|   |   |   |--- EstimatedSalary >  -0.06
|   |   |   |   |--- EstimatedSalary <= 0.40
|   |   |   |   |   |--- EstimatedSalary <= 0.03
|   |   |   |   |   |   |--- class: 1

回答7:

Just because everyone was so helpful I'll just add a modification to Zelazny7 and Daniele's beautiful solutions. This one is for python 2.7, with tabs to make it more readable:

def get_code(tree, feature_names, tabdepth=0):
    left      = tree.tree_.children_left
    right     = tree.tree_.children_right
    threshold = tree.tree_.threshold
    features  = [feature_names[i] for i in tree.tree_.feature]
    value = tree.tree_.value

    def recurse(left, right, threshold, features, node, tabdepth=0):
            if (threshold[node] != -2):
                    print '\t' * tabdepth,
                    print "if ( " + features[node] + " <= " + str(threshold[node]) + " ) {"
                    if left[node] != -1:
                            recurse (left, right, threshold, features,left[node], tabdepth+1)
                    print '\t' * tabdepth,
                    print "} else {"
                    if right[node] != -1:
                            recurse (left, right, threshold, features,right[node], tabdepth+1)
                    print '\t' * tabdepth,
                    print "}"
            else:
                    print '\t' * tabdepth,
                    print "return " + str(value[node])

    recurse(left, right, threshold, features, 0)

回答8:

Codes below is my approach under anaconda python 2.7 plus a package name "pydot-ng" to making a PDF file with decision rules. I hope it is helpful.

from sklearn import tree

clf = tree.DecisionTreeClassifier(max_leaf_nodes=n)
clf_ = clf.fit(X, data_y)

feature_names = X.columns
class_name = clf_.classes_.astype(int).astype(str)

def output_pdf(clf_, name):
    from sklearn import tree
    from sklearn.externals.six import StringIO
    import pydot_ng as pydot
    dot_data = StringIO()
    tree.export_graphviz(clf_, out_file=dot_data,
                         feature_names=feature_names,
                         class_names=class_name,
                         filled=True, rounded=True,
                         special_characters=True,
                          node_ids=1,)
    graph = pydot.graph_from_dot_data(dot_data.getvalue())
    graph.write_pdf("%s.pdf"%name)

output_pdf(clf_, name='filename%s'%n)

a tree graphy show here

回答9:

I've been going through this, but i needed the rules to be written in this format

if A>0.4 then if B<0.2 then if C>0.8 then class='X' 

So I adapted the answer of @paulkernfeld (thanks) that you can customize to your need

def tree_to_code(tree, feature_names, Y):
    tree_ = tree.tree_
    feature_name = [
        feature_names[i] if i != _tree.TREE_UNDEFINED else "undefined!"
        for i in tree_.feature
    ]
    pathto=dict()

    global k
    k = 0
    def recurse(node, depth, parent):
        global k
        indent = "  " * depth

        if tree_.feature[node] != _tree.TREE_UNDEFINED:
            name = feature_name[node]
            threshold = tree_.threshold[node]
            s= "{} <= {} ".format( name, threshold, node )
            if node == 0:
                pathto[node]=s
            else:
                pathto[node]=pathto[parent]+' & ' +s

            recurse(tree_.children_left[node], depth + 1, node)
            s="{} > {}".format( name, threshold)
            if node == 0:
                pathto[node]=s
            else:
                pathto[node]=pathto[parent]+' & ' +s
            recurse(tree_.children_right[node], depth + 1, node)
        else:
            k=k+1
            print(k,')',pathto[parent], tree_.value[node])
    recurse(0, 1, 0)

回答10:

This builds on @paulkernfeld 's answer. If you have a dataframe X with your features and a target dataframe y with your resonses and you you want to get an idea which y value ended in which node (and also ant to plot it accordingly) you can do the following:

    def tree_to_code(tree, feature_names):
        from sklearn.tree import _tree
        codelines = []
        codelines.append('def get_cat(X_tmp):\n')
        codelines.append('   catout = []\n')
        codelines.append('   for codelines in range(0,X_tmp.shape[0]):\n')
        codelines.append('      Xin = X_tmp.iloc[codelines]\n')
        tree_ = tree.tree_
        feature_name = [
            feature_names[i] if i != _tree.TREE_UNDEFINED else "undefined!"
            for i in tree_.feature
        ]
        #print "def tree({}):".format(", ".join(feature_names))

        def recurse(node, depth):
            indent = "      " * depth
            if tree_.feature[node] != _tree.TREE_UNDEFINED:
                name = feature_name[node]
                threshold = tree_.threshold[node]
                codelines.append ('{}if Xin["{}"] <= {}:\n'.format(indent, name, threshold))
                recurse(tree_.children_left[node], depth + 1)
                codelines.append( '{}else:  # if Xin["{}"] > {}\n'.format(indent, name, threshold))
                recurse(tree_.children_right[node], depth + 1)
            else:
                codelines.append( '{}mycat = {}\n'.format(indent, node))

        recurse(0, 1)
        codelines.append('      catout.append(mycat)\n')
        codelines.append('   return pd.DataFrame(catout,index=X_tmp.index,columns=["category"])\n')
        codelines.append('node_ids = get_cat(X)\n')
        return codelines
    mycode = tree_to_code(clf,X.columns.values)

    # now execute the function and obtain the dataframe with all nodes
    exec(''.join(mycode))
    node_ids = [int(x[0]) for x in node_ids.values]
    node_ids2 = pd.DataFrame(node_ids)

    print('make plot')
    import matplotlib.cm as cm
    colors = cm.rainbow(np.linspace(0, 1, 1+max( list(set(node_ids)))))
    #plt.figure(figsize=cm2inch(24, 21))
    for i in list(set(node_ids)):
        plt.plot(y[node_ids2.values==i],'o',color=colors[i], label=str(i))  
    mytitle = ['y colored by node']
    plt.title(mytitle ,fontsize=14)
    plt.xlabel('my xlabel')
    plt.ylabel(tagname)
    plt.xticks(rotation=70)       
    plt.legend(loc='upper center', bbox_to_anchor=(0.5, 1.00), shadow=True, ncol=9)
    plt.tight_layout()
    plt.show()
    plt.close 

not the most elegant version but it does the job...

回答11:

This is the code you need

I have modified the top liked code to indent in a jupyter notebook python 3 correctly

import numpy as np
from sklearn.tree import _tree

def tree_to_code(tree, feature_names):
    tree_ = tree.tree_
    feature_name = [feature_names[i] 
                    if i != _tree.TREE_UNDEFINED else "undefined!" 
                    for i in tree_.feature]
    print("def tree({}):".format(", ".join(feature_names)))

    def recurse(node, depth):
        indent = "    " * depth
        if tree_.feature[node] != _tree.TREE_UNDEFINED:
            name = feature_name[node]
            threshold = tree_.threshold[node]
            print("{}if {} <= {}:".format(indent, name, threshold))
            recurse(tree_.children_left[node], depth + 1)
            print("{}else:  # if {} > {}".format(indent, name, threshold))
            recurse(tree_.children_right[node], depth + 1)
        else:
            print("{}return {}".format(indent, np.argmax(tree_.value[node])))

    recurse(0, 1)

回答12:

Here is a function, printing rules of a scikit-learn decision tree under python 3 and with offsets for conditional blocks to make the structure more readable:

def print_decision_tree(tree, feature_names=None, offset_unit='    '):
    '''Plots textual representation of rules of a decision tree
    tree: scikit-learn representation of tree
    feature_names: list of feature names. They are set to f1,f2,f3,... if not specified
    offset_unit: a string of offset of the conditional block'''

    left      = tree.tree_.children_left
    right     = tree.tree_.children_right
    threshold = tree.tree_.threshold
    value = tree.tree_.value
    if feature_names is None:
        features  = ['f%d'%i for i in tree.tree_.feature]
    else:
        features  = [feature_names[i] for i in tree.tree_.feature]        

    def recurse(left, right, threshold, features, node, depth=0):
            offset = offset_unit*depth
            if (threshold[node] != -2):
                    print(offset+"if ( " + features[node] + " <= " + str(threshold[node]) + " ) {")
                    if left[node] != -1:
                            recurse (left, right, threshold, features,left[node],depth+1)
                    print(offset+"} else {")
                    if right[node] != -1:
                            recurse (left, right, threshold, features,right[node],depth+1)
                    print(offset+"}")
            else:
                    print(offset+"return " + str(value[node]))

    recurse(left, right, threshold, features, 0,0)

回答13:

Here is a way to translate the whole tree into a single (not necessarily too human-readable) python expression using the SKompiler library:

from skompiler import skompile
skompile(dtree.predict).to('python/code')

回答14:

You can also make it more informative by distinguishing it to which class it belongs or even by mentioning its output value.

def print_decision_tree(tree, feature_names, offset_unit='    '):    
left      = tree.tree_.children_left
right     = tree.tree_.children_right
threshold = tree.tree_.threshold
value = tree.tree_.value
if feature_names is None:
    features  = ['f%d'%i for i in tree.tree_.feature]
else:
    features  = [feature_names[i] for i in tree.tree_.feature]        

def recurse(left, right, threshold, features, node, depth=0):
        offset = offset_unit*depth
        if (threshold[node] != -2):
                print(offset+"if ( " + features[node] + " <= " + str(threshold[node]) + " ) {")
                if left[node] != -1:
                        recurse (left, right, threshold, features,left[node],depth+1)
                print(offset+"} else {")
                if right[node] != -1:
                        recurse (left, right, threshold, features,right[node],depth+1)
                print(offset+"}")
        else:
                #print(offset,value[node]) 

                #To remove values from node
                temp=str(value[node])
                mid=len(temp)//2
                tempx=[]
                tempy=[]
                cnt=0
                for i in temp:
                    if cnt<=mid:
                        tempx.append(i)
                        cnt+=1
                    else:
                        tempy.append(i)
                        cnt+=1
                val_yes=[]
                val_no=[]
                res=[]
                for j in tempx:
                    if j=="[" or j=="]" or j=="." or j==" ":
                        res.append(j)
                    else:
                        val_no.append(j)
                for j in tempy:
                    if j=="[" or j=="]" or j=="." or j==" ":
                        res.append(j)
                    else:
                        val_yes.append(j)
                val_yes = int("".join(map(str, val_yes)))
                val_no = int("".join(map(str, val_no)))

                if val_yes>val_no:
                    print(offset,'\033[1m',"YES")
                    print('\033[0m')
                elif val_no>val_yes:
                    print(offset,'\033[1m',"NO")
                    print('\033[0m')
                else:
                    print(offset,'\033[1m',"Tie")
                    print('\033[0m')

recurse(left, right, threshold, features, 0,0)

回答15:

Here is my approach to extract the decision rules in a form that can be used in directly in sql, so the data can be grouped by node. (Based on the approaches of previous posters.)

The result will be subsequent CASE clauses that can be copied to an sql statement, ex.

SELECT COALESCE(*CASE WHEN <conditions> THEN > <NodeA>*, > *CASE WHEN <conditions> THEN <NodeB>*, > ....)NodeName,* > FROM <table or view>

---

import numpy as np

import pickle
feature_names=.............
features  = [feature_names[i] for i in range(len(feature_names))]
clf= pickle.loads(trained_model)
impurity=clf.tree_.impurity
importances = clf.feature_importances_
SqlOut=""

#global Conts
global ContsNode
global Path
#Conts=[]#
ContsNode=[]
Path=[]
global Results
Results=[]

def print_decision_tree(tree, feature_names, offset_unit=''    ''):    
    left      = tree.tree_.children_left
    right     = tree.tree_.children_right
    threshold = tree.tree_.threshold
    value = tree.tree_.value

    if feature_names is None:
        features  = [''f%d''%i for i in tree.tree_.feature]
    else:
        features  = [feature_names[i] for i in tree.tree_.feature]        

    def recurse(left, right, threshold, features, node, depth=0,ParentNode=0,IsElse=0):
        global Conts
        global ContsNode
        global Path
        global Results
        global LeftParents
        LeftParents=[]
        global RightParents
        RightParents=[]
        for i in range(len(left)): # This is just to tell you how to create a list.
            LeftParents.append(-1)
            RightParents.append(-1)
            ContsNode.append("")
            Path.append("")


        for i in range(len(left)): # i is node
            if (left[i]==-1 and right[i]==-1):      
                if LeftParents[i]>=0:
                    if Path[LeftParents[i]]>" ":
                        Path[i]=Path[LeftParents[i]]+" AND " +ContsNode[LeftParents[i]]                                 
                    else:
                        Path[i]=ContsNode[LeftParents[i]]                                   
                if RightParents[i]>=0:
                    if Path[RightParents[i]]>" ":
                        Path[i]=Path[RightParents[i]]+" AND not " +ContsNode[RightParents[i]]                                   
                    else:
                        Path[i]=" not " +ContsNode[RightParents[i]]                     
                Results.append(" case when  " +Path[i]+"  then ''" +"{:4d}".format(i)+ " "+"{:2.2f}".format(impurity[i])+" "+Path[i][0:180]+"''")

            else:       
                if LeftParents[i]>=0:
                    if Path[LeftParents[i]]>" ":
                        Path[i]=Path[LeftParents[i]]+" AND " +ContsNode[LeftParents[i]]                                 
                    else:
                        Path[i]=ContsNode[LeftParents[i]]                                   
                if RightParents[i]>=0:
                    if Path[RightParents[i]]>" ":
                        Path[i]=Path[RightParents[i]]+" AND not " +ContsNode[RightParents[i]]                                   
                    else:
                        Path[i]=" not "+ContsNode[RightParents[i]]                      
                if (left[i]!=-1):
                    LeftParents[left[i]]=i
                if (right[i]!=-1):
                    RightParents[right[i]]=i
                ContsNode[i]=   "( "+ features[i] + " <= " + str(threshold[i])   + " ) "

    recurse(left, right, threshold, features, 0,0,0,0)
print_decision_tree(clf,features)
SqlOut=""
for i in range(len(Results)): 
    SqlOut=SqlOut+Results[i]+ " end,"+chr(13)+chr(10)

回答16:

Modified Zelazny7's code to fetch SQL from the decision tree.

# SQL from decision tree

def get_lineage(tree, feature_names):
     left      = tree.tree_.children_left
     right     = tree.tree_.children_right
     threshold = tree.tree_.threshold
     features  = [feature_names[i] for i in tree.tree_.feature]
     le='<='               
     g ='>'
     # get ids of child nodes
     idx = np.argwhere(left == -1)[:,0]     

     def recurse(left, right, child, lineage=None):          
          if lineage is None:
               lineage = [child]
          if child in left:
               parent = np.where(left == child)[0].item()
               split = 'l'
          else:
               parent = np.where(right == child)[0].item()
               split = 'r'
          lineage.append((parent, split, threshold[parent], features[parent]))
          if parent == 0:
               lineage.reverse()
               return lineage
          else:
               return recurse(left, right, parent, lineage)
     print 'case '
     for j,child in enumerate(idx):
        clause=' when '
        for node in recurse(left, right, child):
            if len(str(node))<3:
                continue
            i=node
            if i[1]=='l':  sign=le 
            else: sign=g
            clause=clause+i[3]+sign+str(i[2])+' and '
        clause=clause[:-4]+' then '+str(j)
        print clause
     print 'else 99 end as clusters'

回答17:

Apparently a long time ago somebody already decided to try to add the following function to the official scikit's tree export functions (which basically only supports export_graphviz)

def export_dict(tree, feature_names=None, max_depth=None) :
    """Export a decision tree in dict format.

Here is his full commit:

https://github.com/scikit-learn/scikit-learn/blob/79bdc8f711d0af225ed6be9fdb708cea9f98a910/sklearn/tree/export.py

Not exactly sure what happened to this comment. But you could also try to use that function.

I think this warrants a serious documentation request to the good people of scikit-learn to properly document the sklearn.tree.Tree API which is the underlying tree structure that DecisionTreeClassifier exposes as its attribute tree_.

回答18:

Just use the function from sklearn.tree like this

from sklearn.tree import export_graphviz
    export_graphviz(tree,
                out_file = "tree.dot",
                feature_names = tree.columns) //or just ["petal length", "petal width"]

And then look in your project folder for the file tree.dot, copy the ALL the content and paste it here http://www.webgraphviz.com/ and generate your graph :)