Decision Tree Parameters

There are various parameters in a decision tree, and the results change depending on how they are specified. In this page, we will try to visualize and check how each parameter works.

max_depth specifies the maximum depth of the tree
min_samples_split specifies the minimum number of data required to create a branch.
min_samples_leaf specifies the minimum number of data required to create a leaf.
max_leaf_nodes specifies the maximum number of leaves.
ccp_alpha is a parameter for pruning the decision tree to account for tree complexity
class_weight specifies the weighting of classes in classification.

import matplotlib.pyplot as plt
from sklearn.tree import DecisionTreeRegressor
from sklearn.datasets import make_regression
from mpl_toolkits.mplot3d import Axes3D
from dtreeviz.trees import dtreeviz, rtreeviz_bivar_3D

Applying a decision tree to simple dataset #

# dataset
X, y = make_regression(n_samples=100, n_features=2, random_state=11)

# train decision tree
dt = DecisionTreeRegressor(max_depth=3)
dt.fit(X, y)

# visualize
fig = plt.figure(figsize=(10, 10))
ax = fig.add_subplot(111, projection="3d")
t = rtreeviz_bivar_3D(
    dt,
    X,
    y,
    feature_names=["x1", "x2"],
    target_name="MPG",
    elev=40,
    azim=120,
    dist=8.0,
    show={"splits", "title"},
    ax=ax,
)
plt.show()

png

Mempelajari pohon keputusan dengan berbagai parameter #

Mari kita periksa bagaimana pohon keputusan dengan struktur yang sedikit kompleks berperilaku ketika parameter pohon keputusan diubah. Pertama, periksa pohon keputusan dengan nilai default untuk semua parameter kecuali max_depth=3.

X, y = make_regression(
    n_samples=500, n_features=2, effective_rank=4, noise=0.1, random_state=1
)
plt.figure(figsize=(10, 10))
plt.scatter(X[:, 0], X[:, 1], c=y)
plt.show()

dt = DecisionTreeRegressor(max_depth=3, random_state=117117)
dt.fit(X, y)

fig = plt.figure(figsize=(10, 10))
ax = fig.add_subplot(111, projection="3d")
t = rtreeviz_bivar_3D(
    dt,
    X,
    y,
    feature_names=["x1", "x2"],
    target_name="y",
    elev=40,
    azim=240,
    dist=8.0,
    show={"splits", "title"},
    ax=ax,
)
plt.show()

png

max_depth = 10 #

When the value of max_depth is large, a deeper and more complex tree is created. This can represent complex rules, but may be over-fitting if the number of data is small.

dt = DecisionTreeRegressor(max_depth=10, random_state=117117)
dt.fit(X, y)

fig = plt.figure(figsize=(10, 10))
ax = fig.add_subplot(111, projection="3d")
t = rtreeviz_bivar_3D(
    dt,
    X,
    y,
    feature_names=["x1", "x2"],
    target_name="y",
    elev=40,
    azim=240,
    dist=8.0,
    show={"splits", "title"},
    ax=ax,
)
plt.show()

png

max-depth=5 #

dt = DecisionTreeRegressor(max_depth=5, random_state=117117)
dt.fit(X, y)

fig = plt.figure(figsize=(10, 10))
ax = fig.add_subplot(111, projection="3d")
t = rtreeviz_bivar_3D(
    dt,
    X,
    y,
    feature_names=["x1", "x2"],
    target_name="y",
    elev=40,
    azim=240,
    dist=8.0,
    show={"splits", "title"},
    ax=ax,
)
plt.show()

png

min_samples_split=60 #

Specifies the minimum number of data required to create a single split. Smaller numbers of min_samples_split allow for more detailed rules. If you increase the number, you can avoid over-fittinging.

dt = DecisionTreeRegressor(max_depth=5, min_samples_split=60, random_state=117117)
dt.fit(X, y)

fig = plt.figure(figsize=(10, 10))
ax = fig.add_subplot(111, projection="3d")
t = rtreeviz_bivar_3D(
    dt,
    X,
    y,
    feature_names=["x1", "x2"],
    target_name="y",
    elev=40,
    azim=240,
    dist=8.0,
    show={"splits", "title"},
    ax=ax,
)
plt.show()

png

ccp_alpha=0.4 #

This parameter penalizes the complexity of the tree. The higher the value of ccp_alpha, the simpler the tree will be.

dt = DecisionTreeRegressor(max_depth=5, random_state=117117, ccp_alpha=0.4)
dt.fit(X, y)

fig = plt.figure(figsize=(10, 10))
ax = fig.add_subplot(111, projection="3d")
t = rtreeviz_bivar_3D(
    dt,
    X,
    y,
    feature_names=["x1", "x2"],
    target_name="y",
    elev=40,
    azim=240,
    dist=8.0,
    show={"splits", "title"},
    ax=ax,
)
plt.show()

png

max_leaf_nodes=5 #

This parameter specifies the number of leaves that will eventually be created. The number of max_leaf_nodes matches the number of parcels.

dt = DecisionTreeRegressor(max_depth=5, random_state=117117, max_leaf_nodes=5)
dt.fit(X, y)

fig = plt.figure(figsize=(10, 10))
ax = fig.add_subplot(111, projection="3d")
t = rtreeviz_bivar_3D(
    dt,
    X,
    y,
    feature_names=["x1", "x2"],
    target_name="y",
    elev=40,
    azim=240,
    dist=8.0,
    show={"splits", "title"},
    ax=ax,
)
plt.show()

png

when dataset contain outliers #

Specify which criterion to apply when creating a branch. Let’s see how the tree changes when criterion="squared_error" is specified with outliers. Since squared_error penalizes outliers more strongly than absolute_error, it is expected that a decision tree branch will be created if squared_error is specified.

## Multiply some data values by 5 as outlier
X, y = make_regression(n_samples=100, n_features=2, random_state=11)
y[1:20] = y[1:20] * 5

dt = DecisionTreeRegressor(max_depth=5, random_state=117117, criterion="absolute_error")
dt.fit(X, y)

fig = plt.figure(figsize=(10, 10))
ax = fig.add_subplot(111, projection="3d")
t = rtreeviz_bivar_3D(
    dt,
    X,
    y,
    feature_names=["x1", "x2"],
    target_name="y",
    elev=40,
    azim=240,
    dist=8.0,
    show={"splits", "title"},
    ax=ax,
)
plt.show()

png

dt = DecisionTreeRegressor(max_depth=5, random_state=117117, criterion="squared_error")

dt.fit(X, y)

fig = plt.figure(figsize=(10, 10))
ax = fig.add_subplot(111, projection="3d")
t = rtreeviz_bivar_3D(
    dt,
    X,
    y,
    feature_names=["x1", "x2"],
    target_name="y",
    elev=40,
    azim=240,
    dist=8.0,
    show={"splits", "title"},
    ax=ax,
)
plt.show()

png