🛠️ Installation

To install this package clone the repository and run

cd PyFD/
python3 -m pip install .

🔍 Unification of Additive Explanations

The typical Machine Learning methodology is as follows

from pyfd.data import get_data_bike
x, y, features = get_data_bike()
x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.2, random_state=42)
model = HistGradientBoostingRegressor(random_state=0).fit(x_train, y_train)
print(model.score(x_test, y_test))

You load a dataset, train a model on one subset of this data, and evaluate its performance on the other subset. But is reporting performance on held-off data enough to assess that the model is trustworthy? Not necessarily. It would be preferable to get some insight into the model behavior to be sure that it makes correct decisions for the right reason. To get insights into model behavior, various post-hoc techniques have been developped, notably Partial Dependence Plots (PDP), Permutation Feature Importance (PFI), and SHAP.

The motivation behind the PyFD package was to realize that these various explanability methods can be unified through the lens of Functional Decomposition (Our Paper). More specifically, let $B$ be a probability distribution over the input space. Then the model $h$ can be decomposed as follows

$$ h(x) = E_{z\sim B}[h(z)] + \sum_i h_{i,B}(x_i) + \sum_{i\lt j} h_{ij,B}(x_{ij}) + \ldots,$$

where $h_{i, B}(x_i)$ are called the main effects and only depend on a single feature. The remaining terms $h_{u,B}(x_u)$ with $|u|\geq 2$ are called interactions and depend on multiple features simultaneously. The PDP/SHAP/PFI explainers can all be expressed in terms of this functional decomposition.

This has important practical implications : any package that can efficiently compute this functional decomposition can compute any explanation method. This is why PyFD exists.

from pyfd.decompositions import get_components_tree
from pyfd.shapley import interventional_treeshap

# The background distribution B
background = x_train[:1000]
# Compute the h_1, h_2, ..., terms of the Functional Decomposition
decomposition = get_components_tree(model, background, background, features)
# Compute the Shapley Values
shap_values = interventional_treeshap(model, background, background, features)

The functional decomposition can be leveraged to visualize the model behavior locally.

from pyfd.plots import attrib_scatter_plot

# We plot the h_i terms along side the Shapley Values
attrib_scatter_plot(decomposition, shap_values, background, features)

Or globally.

from pyfd.decompositions import get_PDP_PFI_importance
from pyfd.shapley import get_SHAP_importance
from pyfd.plots import bar

I_PDP, I_PFI = get_PDP_PFI_importance(decomposition)
I_SHAP = get_SHAP_importance(shap_values)
bar([I_PFI, I_SHAP, I_PDP], features.names())

📏 Increasing Explanation Alignment

In the bar chart shown previously, the three methods attributed very different importance to the feature workingday. This is problematic because there is no ground-truth for the optimal explanations and so practitionners are left wondering which explanation should I believe? In PyFD we do not shy away from showing contradicting explanations to the user. In fact, we let disagreements be an incentive for the user to increase the alignment between the techniques.

One of the implications of unifying PDP/SHAP/PFI through the lens of functional decomposition is that the root cause of their disagreements can be identified : feature interactions. Thus, PyFD offers methodologies to minimize feature interactions, one of them being regional explanations with Functional Decomposition Trees (FD-Trees for short).

from pyfd.fd_trees import CoE_Tree

tree = CoE_Tree(max_depth=2, features=features)
tree.fit(background, decomposition)
tree.print()
>>>
# Samples 1000
# If workingday ≤ 0.0000:
# |   If hr ≤ 7.0000:
# |   |   Samples 105
# |   |   Region 0
# |   else:
# |   |   Samples 223
# |   |   Region 1
# else:
# |   If hr ≤ 6.0000:
# |   |   Samples 190
# |   |   Region 2
# |   else:
# |   |   Samples 482
# |   |   Region 3

The FD-Tree has decided to partition the input space into four regions based on the workingday and hr features. We can now compute functional decomposition while restricting the distribution $B$ to each distinct region.

# Assign each datum to its region
regions = tree.predict(background)
# Describe each region with a rule
rules = tree.rules()
# Compute regional explanations
regional_backgrounds = [[], [], [], []]
regional_decomposition = [[], [], [], []]
regional_shap = [[], [], [], []]
for r in range(4):
    regional_backgrounds[r] = background[regions==r]
    # Regional decomposition
    regional_decomposition[r] = get_components_tree(model,
                                                    regional_backgrounds[r],
                                                    regional_backgrounds[r],
                                                    features, anchored=true)
    # Shapley values
    regional_shap[r] = interventional_treeshap(model,
                                               regional_backgrounds[r],
                                               regional_backgrounds[r],
                                               features)

And visualize the local and global model behaviors.

attrib_scatter_plot(regional_decomposition, regional_shap, regional_backgrounds, features)
plot_legend(rules)

fig, axes = plt.subplots(1, 4)
for r in range(4):
    I_PDP, I_PFI = get_PDP_PFI_importance(regional_decomposition[r])
    I_SHAP = get_SHAP_importance(regional_shap[r])
    bar([I_PFI, I_SHAP, I_PDP], features.names(), ax=axes[r])
    axes[r].set_title(rules[r])

As a result of partioning the input space, there is almost perfect agreement between the PDP/SHAP/PFI feature importance. We no longer have to wonder which explanation is the correct one since they all agree!

🧠 Tutorials

The PyFD is explained in-depth here in a series of jupyter notebooks.