def display_shapley(X_train, y_train, cf=["dcor"], xlabels=None):
if not isinstance(cf, list):
cf = list(cf)
d = X_train.shape[1]
x_range = list(range(d))
_, ax = plt.subplots()
ax = nice_axes(ax)
for _n, _cf in enumerate(cf):
print(_cf)
_shapley_values = shapley.calc_shapley_values(X_train, y_train, _cf)
#_shapley_values = normalise(_shapley_values)
plt.bar(x_range + 0.1 * _n * numpy.ones(d),
_shapley_values,
alpha=0.5,
label=_cf,
width=0.1)
#if _n < 1:
#_shapley_values, xlabels = [list(t) for t in zip(*sorted(zip(_shapley_values, xlabels), reverse=True))]
if xlabels is None:
x_labels = x_range
ax.set_xticks(x_range)
ax.set_xticklabels(xlabels, rotation=90)
plt.title(r"Shapley decomposition of {0} on training data".format(cf))
plt.legend()
plt.draw()
def display_shapley_vs_xgb(X_test, y_test, y_pred, cf="dcor"):
d = X_test.shape[1]
shapley_values_actual = shapley.calc_shapley_values(X_test, y_test, cf)
shapley_values_xgb = shapley.calc_shapley_values(X_test, y_pred, cf)
_, ax = plt.subplots()
ax = nice_axes(ax)
plt.title(
r"Shapley decomposition of {0} on $X$ with true vs predicted $Y$".
format(cf))
plt.bar(range(len(shapley_values_actual)),
shapley_values_actual,
color="red",
alpha=0.5,
label="True")
plt.bar(range(len(shapley_values_xgb)),
shapley_values_xgb,
color="blue",
alpha=0.5,
label="Predicted")
plt.legend()
plt.draw()
def display_residuals_shapley(x, residuals, cf="dcor"):
d = x.shape[1]
shapley_values_residuals = shapley.calc_shapley_values(x, residuals, cf)
_, ax = plt.subplots()
ax = shapley.nice_axes(ax)
plt.bar(range(len(shapley_values_residuals)),
shapley_values_residuals,
color="red",
alpha=0.5)
plt.title(
r"Shapley decomposition of {0} on $X$ with the residuals (true-predicted $Y$)"
.format(cf))
plt.draw()
sigma = 0.2
X = numpy.array([numpy.random.uniform(-1, 1, N) for _ in range(D)]).T
X[:, 1] = X[:, 0] + numpy.random.normal(0, sigma, N)
X[:, 2] = X[:, 0] + numpy.random.normal(0, sigma, N)
Y = numpy.matmul(numpy.multiply(X, X), numpy.ones(D))
# ---
X_train, X_test, y_train, y_test = train_test_split(X,
Y,
test_size=0.2,
random_state=7)
# --- Fit model
model = XGBRegressor()
model.fit(X_train, y_train)
# --- Predict
y_pred = model.predict(X_test)
# --- Feature importances
feature_importance = model.feature_importances_
shapley_values_actual = shapley.calc_shapley_values(X_test, y_test, "dcor")
shapley_values_xgb = shapley.calc_shapley_values(X_test, y_pred, "dcor")
print(shapley_values_actual)
print(shapley_values_xgb)
display_predictions()
display_feature_importances()
display_shapley()
# --- Check for NaNs:
X_data = X_data.dropna(axis=0)
#print(X_data.isnull().any())
X_data = np.array(X_data[feats])
y_data = np.array(y_data)
y_data = y_data.T[0]
print(X_data.shape)
print(y_data.shape)
# --- Plot all data
#sns.pairplot(data_cont, hue='sex', size=2.5)
#plt.show()
x_range = list(range(X_data.shape[1]))
shapley_values = shapley.calc_shapley_values(X_data, y_data, x_range, "r2")
print(shapley_values)
shapley_values = shapley.calc_shapley_values(X_data, y_data, x_range, "dcor")
print(shapley_values)
display_shapley(X_data, y_data, cf=["r2"])
display_shapley(X_data, y_data, cf=["dcor"])
#plt.show()
#display_shapley(X_data, y_data, cf=["dcor", "aidc", "r2", "hsic"])
# NOTES
sys.exit()
data = pd.read_csv('processed.cleveland.data', sep=",", header=0)
data = data.iloc[:, [0, 1, 3, 4, 7, 9, 13]]
target = "num"
y = data[target].astype('int')
X = data.drop([target], axis=1)
_, ax = plt.subplots()
if np.isnan(X_shapley).any():
print("Data contains nan. Exiting")
sys.exit()
for _n, _cf in enumerate(["dcor", "r2", "aidc"]):
print(_cf)
_sfilename = "results/shapley_expl_{0}_{1}.pickle".format(
_cf, modelname)
if os.path.isfile(_sfilename):
with open(_sfilename, 'rb') as _f:
_shapley_values = pickle.load(_f)
else:
_shapley_values = shapley.calc_shapley_values(
X_shapley, y_shapley, x_range, _cf)
with open(_sfilename, 'wb') as _f:
pickle.dump(_shapley_values, _f)
print(_shapley_values)
plt.bar(x_range + 0.1 * _n * np.ones(d),
_shapley_values,
alpha=0.5,
label=_cf,
width=0.1)
plt.title("Target")
plt.legend()
ax.set_xticks(x_range)
ax.set_xticklabels(labels, rotation=90)
plt.draw()
2.6854023615295426, -0.7195432425919883, 0.4681875927260907,
0.027206421899317337, 0.6396687066390925, 0.203168750153398,
-1.0873217125552856
])
def aidc(x, y):
cov_y = np.cov(y)
cov_x = np.cov(x.T)
if cov_x.shape is ():
inv_cov_x = 1.0 / cov_x
x_trans = np.dot(x, np.sqrt(inv_cov_x))
else:
inv_cov_x = np.linalg.inv(cov_x)
x_trans = np.dot(x, scipy.linalg.sqrtm(inv_cov_x))
inv_cov_y = 1 / cov_y
y_trans = np.dot(y, np.sqrt(inv_cov_y))
return dcor.distance_correlation(x_trans, y_trans)
#print(aidc(x, y))
#print(aidc(x1, y1))
#print(dcor.distance_correlation_af_inv(x, y))
#print(dcor.distance_correlation_af_inv(x1, y1))
#print(calc_shapley_values(np.array(x), np.array(y), cf_name="r2"))
#print(calc_shapley_values(np.array(x), np.array(y), cf_name="dcor"))
print(calc_shapley_values(np.array(x), np.array(y), cf_name="aidc"))
def do_shapley(modelname, preds, labels):
_sfilename = f"results/shapley_features_{modelname}.pickle"
if not os.path.isfile(_sfilename):
with open(_sfilename, 'wb') as _f:
pickle.dump(labels, _f)
# --- On data
d = X_shapley.shape[1]
x_range = list(range(d))
_, ax = plt.subplots()
if np.isnan(X_shapley).any():
print("Data contains nan. Exiting")
sys.exit()
for _n, _cf in enumerate(["dcor", "r2", "aidc"]):
print(_cf)
_sfilename = "results/shapley_expl_{0}_{1}.pickle".format(
_cf, modelname)
if os.path.isfile(_sfilename):
with open(_sfilename, 'rb') as _f:
_shapley_values = pickle.load(_f)
else:
_shapley_values = shapley.calc_shapley_values(
X_shapley, y_shapley, x_range, _cf)
with open(_sfilename, 'wb') as _f:
pickle.dump(_shapley_values, _f)
print(_shapley_values)
plt.bar(x_range + 0.1 * _n * np.ones(d),
_shapley_values,
alpha=0.5,
label=_cf,
width=0.1)
plt.title("Target")
plt.legend()
ax.set_xticks(x_range)
ax.set_xticklabels(labels, rotation=90)
plt.draw()
# --- On predictions
X_shapley_pred = X_test.copy()
X_shapley_pred["sex_isFemale"] = [
1 if _x else 0 for _x in X_shapley_pred["sex_isFemale"]
]
labels = X_shapley_pred.columns
X_shapley_pred = np.array(X_shapley_pred[shapley_features])
_, ax = plt.subplots()
if np.isnan(X_shapley_pred).any():
print("Data contains nan. Exiting")
sys.exit()
for _n, _cf in enumerate(["dcor", "r2", "aidc"]):
print(_cf)
_sfilename = "results/shapley_pred_{0}_{1}.pickle".format(
_cf, modelname)
if os.path.isfile(_sfilename):
with open(_sfilename, 'rb') as _f:
_shapley_values = pickle.load(_f)
else:
_shapley_values = shapley.calc_shapley_values(
X_shapley_pred, preds, x_range, _cf)
with open(_sfilename, 'wb') as _f:
pickle.dump(_shapley_values, _f)
print(_shapley_values)
plt.bar(x_range + 0.1 * _n * np.ones(d),
_shapley_values,
alpha=0.5,
label=_cf,
width=0.1)
plt.title("Predictions")
plt.legend()
ax.set_xticks(x_range)
ax.set_xticklabels(labels, rotation=90)
plt.draw()
# --- On residuals
_, ax = plt.subplots()
residuals = y_test - preds
print(X_shapley_pred.shape)
print(residuals.shape)
if np.isnan(X_shapley_pred).any():
print("Data contains nan. Exiting")
sys.exit()
for _n, _cf in enumerate(["dcor", "r2", "aidc"]):
print(_cf)
_sfilename = "results/shapley_res_{0}_{1}.pickle".format(
_cf, modelname)
if os.path.isfile(_sfilename):
with open(_sfilename, 'rb') as _f:
_shapley_values = pickle.load(_f)
else:
_shapley_values = shapley.calc_shapley_values(
X_shapley_pred, residuals, x_range, _cf)
with open(_sfilename, 'wb') as _f:
pickle.dump(_shapley_values, _f)
print(_shapley_values)
plt.bar(x_range + 0.1 * _n * np.ones(d),
_shapley_values,
alpha=0.5,
label=_cf,
width=0.1)
plt.title("Residuals")
plt.legend()
ax.set_xticks(x_range)
ax.set_xticklabels(labels, rotation=90)
plt.draw()