決定木の可視化と解釈
sklearnのplot_treeで木を出力・描画し、決定ルールを読み取って、関係者向けレポート用に特徴量の重要度を抽出します。
「決定木の可視化と解釈」はCoddyKit上の無料Machine Learning Academyレッスンです。 これはレッスン4/4です。 下記で完全なレッスンを無料で読むことができます。その後、ブラウザ内の組み込みコードエディタと24時間対応のAIチューターでハンズオン演習できます。 これはMachine Learning Academy学習パスの一部であり、ウェブとCoddyKitアプリ全体で進捗が同期されます。 Machine Learning Academyコースには全4レッスンが含まれています。
このレッスンの一部はまだ翻訳されておらず、英語で表示されています。
Why Tree Visualisation Matters
Decision trees are often called white-box models because their decision logic is fully transparent. Visualising a trained tree lets you: verify the model is making decisions based on sensible features, explain predictions to non-technical stakeholders, identify potential data quality issues (e.g., a feature that should not be important appearing at the root), and debug unexpected behaviour. Visualisation turns the tree's mathematical structure into a human-readable flowchart that domain experts can validate against their knowledge.
from sklearn.tree import DecisionTreeClassifier, plot_tree
from sklearn.datasets import load_iris
import matplotlib.pyplot as plt
X, y = load_iris(return_X_y=True)
feature_names = load_iris().feature_names
class_names = load_iris().target_names
tree = DecisionTreeClassifier(max_depth=3, random_state=42)
tree.fit(X, y)
plt.figure(figsize=(14, 6))
plot_tree(tree,
feature_names=feature_names,
class_names=class_names,
filled=True, # Color by majority class
rounded=True, # Rounded boxes
fontsize=10)
plt.title('Iris Decision Tree (depth=3)')
plt.show()Reading a Node in plot_tree Output
Each node in the plot_tree output shows four pieces of information: (1) The split condition (e.g., petal length <= 2.45), (2) The Gini impurity of the node, (3) The number of samples that reached this node during training, and (4) The class distribution as a list of sample counts per class. Leaf nodes show all four but no split condition — the majority class is the prediction. Node colour intensity indicates purity: darker = more samples of the dominant class.
# Interpreting node output from plot_tree:
#
# petal length (cm) <= 2.45 <- split condition
# gini = 0.667 <- impurity before split
# samples = 150 <- training samples reaching node
# value = [50, 50, 50] <- samples per class [setosa, versicolor, virginica]
# class = setosa <- majority class (prediction if leaf)
print('Gini 0.667 = equal 3-class split (maximum 3-class impurity)')
print('samples=150 at root = all training samples')
print('value=[50,50,50] = perfectly balanced classes')Exporting Tree as Text with export_text
For logging, reports, or environments without graphical display, export_text() produces a text-based representation of the tree. Each level of indentation represents one split level. The pipe character shows branches, and leaf lines show the predicted class. This format is useful for embedding decision rules in documentation, saving to log files, or displaying in command-line environments. It also allows comparing tree structures numerically across different hyperparameter configurations.
from sklearn.tree import DecisionTreeClassifier, export_text
from sklearn.datasets import load_iris
X, y = load_iris(return_X_y=True)
tree = DecisionTreeClassifier(max_depth=3, random_state=42)
tree.fit(X, y)
text_repr = export_text(
tree,
feature_names=list(load_iris().feature_names)
)
print(text_repr)Exporting to Graphviz DOT Format
export_graphviz() generates a DOT language file that can be rendered as a high-quality SVG or PNG using Graphviz. This is ideal for presentation-quality tree diagrams and for large trees that need scrolling to view. The DOT file can also be converted to a PDF or embedded in reports. In Jupyter, use graphviz.Source(dot_data) to render inline. This approach gives full control over font size, colour scheme, and layout — important when the tree is shared with business stakeholders.
from sklearn.tree import export_graphviz
from sklearn.datasets import load_iris
from sklearn.tree import DecisionTreeClassifier
import graphviz
X, y = load_iris(return_X_y=True)
tree = DecisionTreeClassifier(max_depth=3, random_state=42)
tree.fit(X, y)
dot_data = export_graphviz(
tree,
out_file=None,
feature_names=load_iris().feature_names,
class_names=load_iris().target_names,
filled=True, rounded=True,
special_characters=True
)
# graph = graphviz.Source(dot_data) # Renders in Jupyter
# graph.render('iris_tree', format='png') # Save as PNGFeature Importances: What Drove the Model?
After training, tree.feature_importances_ reveals the relative contribution of each input feature to the model's predictions. Features used at the root and upper levels typically have high importance because their splits affect all training samples. Features used only in deep leaves have low importance. Plotting feature importances as a bar chart is a standard step in any tree-based analysis — it confirms that the model relies on sensible, domain-relevant features rather than spurious correlates that happen to work on training data.
from sklearn.tree import DecisionTreeClassifier
from sklearn.datasets import load_breast_cancer
import pandas as pd
import matplotlib.pyplot as plt
X, y = load_breast_cancer(return_X_y=True)
features = load_breast_cancer().feature_names
tree = DecisionTreeClassifier(max_depth=5, random_state=42)
tree.fit(X, y)
imp = pd.Series(tree.feature_importances_, index=features).sort_values(ascending=False)
imp.head(10).plot(kind='barh')
plt.title('Top 10 Feature Importances')
plt.xlabel('Importance')
plt.gca().invert_yaxis()
plt.show()
print('Top feature:', imp.index[0], '(importance:', imp.iloc[0].round(3), ')')Tracing a Single Prediction
The decision_path() method returns a sparse indicator matrix showing which nodes each sample visits. Combined with tree.tree_, you can reconstruct the exact sequence of decisions for any prediction. This is the foundation of automated explanation systems: for each prediction, you can generate a human-readable list of rules like 'petal length was 1.4 cm (≤2.45), so went left; ended at leaf predicting setosa.' This level of transparency is required in regulated domains where every decision must be auditable.
from sklearn.tree import DecisionTreeClassifier
from sklearn.datasets import load_iris
import numpy as np
X, y = load_iris(return_X_y=True)
features = load_iris().feature_names
tree = DecisionTreeClassifier(max_depth=3, random_state=42)
tree.fit(X, y)
# Trace the path for sample 0
node_indicator = tree.decision_path(X[[0]])
nodes_visited = node_indicator.indices
T = tree.tree_
for node in nodes_visited[:-1]: # All except leaf
feature = features[T.feature[node]]
threshold = T.threshold[node]
val = X[0, T.feature[node]]
direction = 'left (<= )' if val <= threshold else 'right (> )'
print(f'Node {node}: {feature} = {val:.2f}, threshold={threshold:.2f} -> {direction}')Interpreting Split Conditions for Stakeholders
When communicating tree decisions to non-technical stakeholders, convert the mathematical split conditions into plain-language statements. Instead of 'petal length (cm) <= 2.45: gini=0.0, samples=50', say 'If the petal is shorter than 2.45 cm, the flower is almost certainly a Setosa.' Frame each branch in terms of the business meaning of the feature. Decision trees are uniquely suited to stakeholder communication among ML models because every decision corresponds to a testable, interpretable business rule.
# Human-readable rule extraction from a trained tree
from sklearn.tree import _tree
def extract_rules(tree, feature_names, class_names):
T = tree.tree_
rules = []
def recurse(node, path):
if T.feature[node] != _tree.TREE_UNDEFINED:
feat = feature_names[T.feature[node]]
thresh = T.threshold[node]
recurse(T.children_left[node], path + [f'{feat} <= {thresh:.2f}'])
recurse(T.children_right[node], path + [f'{feat} > {thresh:.2f}'])
else:
majority_class = class_names[T.value[node].argmax()]
rules.append(' AND '.join(path) + f' => {majority_class}')
recurse(0, [])
return rulesPartial Dependence Plots for Individual Features
While feature importances show which features matter most, Partial Dependence Plots (PDP) show how a feature influences the prediction. A PDP marginalises over all other features and plots the model's predicted output as a function of one (or two) features. For decision trees, PDPs produce step-function shapes reflecting the axis-aligned split thresholds. Scikit-learn's PartialDependenceDisplay generates these plots directly from a fitted tree, making it easy to explain individual feature effects to domain experts.
from sklearn.inspection import PartialDependenceDisplay
from sklearn.tree import DecisionTreeClassifier
from sklearn.datasets import load_iris
import matplotlib.pyplot as plt
X, y = load_iris(return_X_y=True)
features = load_iris().feature_names
tree = DecisionTreeClassifier(max_depth=4, random_state=42)
tree.fit(X, y)
# PDP for the two most important features
fig, ax = plt.subplots(figsize=(10, 4))
PartialDependenceDisplay.from_estimator(
tree, X, features=[2, 3], # petal length and petal width
feature_names=features, ax=ax
)
plt.tight_layout()
plt.show()Comparing Tree Structures Across Hyperparameters
Visualising how the tree structure changes with depth helps build intuition. A depth-1 tree (stump) has one split and two leaves — the most important single feature. A depth-2 tree refines both branches with a second level of questions. Comparing trees at depth 1, 3, and 5 on the same dataset shows how the model builds increasingly complex decision logic. If the depth-5 tree uses the same features as the depth-3 tree at its top levels, those features are genuinely important. If new, obscure features appear at depth 5, they are likely capturing noise.
from sklearn.tree import DecisionTreeClassifier, export_text
from sklearn.datasets import load_iris
X, y = load_iris(return_X_y=True)
feat_names = list(load_iris().feature_names)
for depth in [1, 3, 5]:
tree = DecisionTreeClassifier(max_depth=depth, random_state=42)
tree.fit(X, y)
print(f'\n--- max_depth={depth}, leaves={tree.get_n_leaves()} ---')
print(export_text(tree, feature_names=feat_names)[:300])Using Trees to Generate Business Rules
One of the most valuable uses of decision trees in industry is generating explicit business rules that can be implemented in rule engines, spreadsheets, or legacy systems that cannot run ML models. Each path from root to leaf is a complete IF-THEN rule. These rules can be translated into SQL WHERE clauses, Python dictionaries, or credit-scoring scorecards. By carefully controlling tree depth and minimum sample constraints, you can generate a small, high-accuracy rule set that a business analyst can manually review, approve, and maintain.
# Generate SQL-like rules from a trained decision tree
from sklearn.tree import _tree
def tree_to_sql(tree, feature_names, class_names):
T = tree.tree_
rules = []
def traverse(node, conditions):
if T.feature[node] != _tree.TREE_UNDEFINED:
fname = feature_names[T.feature[node]]
thresh = T.threshold[node]
traverse(T.children_left[node],
conditions + [f'{fname} <= {thresh:.3f}'])
traverse(T.children_right[node],
conditions + [f'{fname} > {thresh:.3f}'])
else:
pred = class_names[T.value[node].argmax()]
where = ' AND '.join(conditions)
rules.append(f'WHEN {where} THEN {pred!r}')
traverse(0, [])
return 'CASE\n ' + '\n '.join(rules) + '\nEND',Saving Tree Visualisations to Files
Saving tree visualisations to files makes them shareable in reports, presentations, and model documentation. With plot_tree and matplotlib, save as PNG or SVG using plt.savefig(). With Graphviz, render to PDF directly. For interactive exploration in Jupyter notebooks, inline SVG provides the clearest output because it is infinitely scalable — useful for deep trees that would be blurry as a fixed-resolution PNG. For stakeholder deliverables, always export at high DPI (300+) or as vector SVG format so the text in nodes remains crisp when zoomed.
from sklearn.tree import DecisionTreeClassifier, plot_tree
from sklearn.datasets import load_iris
import matplotlib.pyplot as plt
X, y = load_iris(return_X_y=True)
tree = DecisionTreeClassifier(max_depth=3, random_state=42)
tree.fit(X, y)
# Save as high-DPI PNG for reports
fig, ax = plt.subplots(figsize=(16, 8))
plot_tree(tree, feature_names=load_iris().feature_names,
class_names=load_iris().target_names,
filled=True, rounded=True, ax=ax, fontsize=10)
fig.savefig('iris_decision_tree.png', dpi=200, bbox_inches='tight')
fig.savefig('iris_decision_tree.svg', format='svg', bbox_inches='tight')
print('Saved PNG and SVG tree visualisations')Quick Check
Test your understanding of Machine Learning with Python concepts from this lesson.
Lesson Recap
In this lesson you learned: how to visualise decision trees using plot_tree, export_text, and export_graphviz, how to read node information (split condition, Gini, samples, value), and how to extract feature importances and decision paths for stakeholder communication. Next up we explore Naive Bayes — a probabilistic classifier that applies Bayes' theorem to make predictions.
よくある質問
「決定木の可視化と解釈」レッスンは無料ですか？
はい。「決定木の可視化と解釈」の完全なテキストはこのウェブで無料で読めます。インタラクティブに演習し（組み込みコードエディタと24時間対応のAIチューター）、Machine Learning Academyコースの残りをアンロックするには、CoddyKit PROにアップグレードしてください。 Machine Learning Academyコースには全4レッスンが含まれています。
「決定木の可視化と解釈」で何を学びますか？
sklearnのplot_treeで木を出力・描画し、決定ルールを読み取って、関係者向けレポート用に特徴量の重要度を抽出します。 ブラウザで直接実行するハンズオンコードでMachine Learning Academyを演習し、24時間対応のAIチューターがレッスンを進める中での質問に答えます。
Machine Learning Academyを始めるのに経験は必要ですか？
事前経験は必要ありません。CoddyKitのMachine Learning Academyは初級者から上級者向けに構成されているため、ここから始めるか最初から始めて、自分のペースで進むことができます。 これはレッスン4/4です。
「決定木の可視化と解釈」レッスンにはどのくらい時間がかかりますか？
ほとんどのCoddyKitレッスンは約5～10分かかります。各レッスンはコンパクトでインタラクティブなので、着実に進歩し、ウェブとアプリ全体で正確に前回の場所から再開できます。
このMachine Learning Academyレッスンでコードを書いて実行できますか？
はい。すべてのMachine Learning Academyレッスンに組み込みコードエディタが含まれているため、ブラウザでリアルコードを書いて実行し、即座のAIフィードバックを取得できます。ローカル設定は不要です。
このコースのすべてのレッスン
- 木の構築：分割、ノード、葉
- ジニ不純度と情報利得
- 木の深さを制御して過学習を防ぐ
- 決定木の可視化と解釈