def kfold_cross_validation(alg, X, y, k, valid_threshold=0.9):
"""
:param alg: 算法模型实例
:param X:
:param y:
:param k: kfold算法的数目
:param valid_threshold: 验证通过的阈值(训练集和测试集预测结果的余弦相似度)
:return: mse, auc, threshold
"""
mse_kfold = {"train": [], "test": []}
auc_kfold = {"train": [], "test": []}
threshold_kfold = []
kf = KFold(n_splits=k, shuffle=True)
for idx, (train_idx, test_idx) in enumerate(kf.split(y)):
X_train = X.iloc[train_idx, :]
y_train = y.iloc[train_idx, :]
X_test = X.iloc[test_idx, :]
y_test = y.iloc[test_idx, :]
alg_cp = copy.deepcopy(alg)
alg_cp = alg_cp.fit(X_train, y_train)
y_train_pred = alg_cp.predict(X_train)
y_test_pred = alg_cp.predict(X_test)
mse_train_value = mse(y_train, y_train_pred)
auc_train_value = auc(y_train, y_train_pred)
mse_test_value = mse(y_test, y_test_pred)
auc_test_value = auc(y_test, y_test_pred)
fpr, tpr, threshold = roc_curve(y_test, y_test_pred)
f1score = []
for thr in threshold:
y_test_pred_class = np.apply_along_axis(func1d=lambda x: 1
if x >= thr else 0,
axis=1,
arr=y_test_pred)
f1score.append(f1_score(y_test, y_test_pred_class))
threshold_optimal = threshold[np.argmax(f1score)]
mse_kfold["train"].append(mse_train_value)
mse_kfold["test"].append(mse_test_value)
auc_kfold["train"].append(auc_train_value)
auc_kfold["test"].append(auc_test_value)
threshold_kfold.append(threshold_optimal)
alg.logger.info(
"cross validation fold[%d]: mse_train[%.6f], mse_test[%.6f], auc_train[%.6f], auc_test[%.6f], threshold[%.6f]"
% (idx, mse_train_value, mse_test_value, auc_train_value,
auc_test_value, threshold_optimal))
cos_sim_mse = \
np.dot(mse_kfold["train"], mse_kfold["test"]) / \
(np.linalg.norm(mse_kfold["train"]) * np.linalg.norm(mse_kfold["test"]))
cos_sim_auc = \
np.dot(auc_kfold["train"], auc_kfold["test"]) / \
(np.linalg.norm(auc_kfold["train"]) * np.linalg.norm(auc_kfold["test"]))
alg.logger.info("cross validation: cosine similarity of mse: %.6f" %
cos_sim_mse)
alg.logger.info("cross validation: cosine similarity of auc: %.6f" %
cos_sim_auc)
if cos_sim_mse < valid_threshold or cos_sim_auc < valid_threshold:
alg.logger.error("cross validation: algorithm is overfit")
raise Exception("%s cross validation: algorithm is overfit" %
alg.__class__.__name__)
return np.mean(mse_kfold["test"]), np.mean(
auc_kfold["test"]), np.mean(threshold_kfold)
def fit(self, X, y, *args, **kwargs):
n_row, n_col = X.shape
## convert to scipy.sparse_matrix
X = csr_matrix(X, dtype=np.float64)
y = csr_matrix(y, dtype=np.float64)
x_sts = np.matrix(X.transpose().sum(axis=1))
x_ratio = csr_matrix(
np.divide(1.0, x_sts, out=np.zeros_like(x_sts), where=x_sts != 0))
## init factors
w0 = self._w0_default
w = csr_matrix(np.full((n_col, 1), self._w_default))
## train model
mse_value = 1.0
logloss_value = sys.maxint
auc_value = 0.0
for i in xrange(self._max_iter):
## delta
delta = self.sig_mod(X, w0, w) - y
## descent
descent_w0 = np.sum(delta)
descent_w = X.transpose().dot(delta)
## penalty
if self._penalty == "L2":
penalty_w0 = self._C * w0
penalty_w = self._C * w
elif self._penalty == "L1":
penalty_w0 = self._C
penalty_w = csr_matrix(np.full((n_col, 1), self._C))
else:
raise Exception("FM must have penalty with one of [L1, L2]")
## gradient descent
w0 = w0 - self._step * (descent_w0 / float(n_row) + penalty_w0)
w = w - self._step * (descent_w.multiply(x_ratio) + penalty_w)
## evaluate
if (i + 1) % 10 == 0 or (i + 1) == self._max_iter:
y_true = y.toarray()
y_pred = self.sig_mod(X, w0, w).toarray()
## MSE
mse_value_new = mse(y_true, y_pred)
mse_ratio = (mse_value - mse_value_new) / mse_value
mse_value = mse_value_new
## LogLoss
logloss_value_new = logloss(y_true, y_pred)
logloss_ratio = (logloss_value -
logloss_value_new) / logloss_value
logloss_value = logloss_value_new
## AUC
auc_value = auc(y_true, y_pred)
# self.logger.info("Iterator[%05d]: MSE[%.6f], LogLoss[%.6f], AUC[%.6f]" % (i+1, mse_value, logloss_value, auc_value))
if mse_ratio <= self._tol or logloss_ratio <= self._tol or (
i + 1) == self._max_iter:
self.logger.info(
"Iterator[%05d]: MSE[%.6f], LogLoss[%.6f], AUC[%.6f]" %
(i + 1, mse_value, logloss_value, auc_value))
break
## save model
self._w0 = w0
self._w = w
return self
def fit(self, X, y, *args, **kwargs):
n_row, n_col = X.shape
w0 = tf.Variable([self._w0_default], dtype="float32")
w = tf.Variable(np.full((n_col, 1), self._w_default), dtype="float32")
v = tf.Variable(np.random.randn(n_col, self._v_length), dtype="float32")
if self._penalty == "L2":
penalty_w0 = self._C * tf.pow(w0, 2)
penalty_w = self._C * tf.pow(w, 2)
penalty_v = self._C * tf.pow(v, 2)
elif self._penalty == "L1":
penalty_w0 = tf.Variable([self._C], dtype="float32")
penalty_w = tf.Variable(np.full((n_col, 1), self._C), dtype="float32")
penalty_v = tf.Variable(np.full((n_col, self._v_length), self._C), dtype="float32")
else:
raise Exception("FM must have penalty with one of [L1, L2]")
xs = tf.constant(np.matrix(X), dtype="float32")
ys = tf.constant(np.matrix(y), dtype="float32")
linear = tf.add(w0, tf.matmul(xs, w))
nonlinear = tf.reduce_sum(tf.pow(tf.matmul(xs, v), 2) - tf.matmul(tf.pow(xs, 2), tf.pow(v, 2)), axis=1, keep_dims=True)
y_pred = 1. / (1. + tf.exp(-tf.add(linear, 0.5 * nonlinear)))
# log_loss = -(ys * tf.log(y_pred) + (1. - ys) * tf.log(1. - y_pred))
square_loss = tf.pow(tf.subtract(ys, y_pred), 2)
loss = tf.reduce_mean(square_loss) + tf.reduce_mean(penalty_w0) + tf.reduce_mean(penalty_w) + tf.reduce_mean(penalty_v)
train_step = tf.train.GradientDescentOptimizer(self._step).minimize(loss)
init = tf.initialize_all_variables()
with tf.Session() as sess:
sess.run(init)
loss_value = sys.maxint
mse_value = 1.0
auc_value = 0.0
for i in range(self._max_iter):
# for i in range(10):
ys_eval = ys.eval()
y_pred_eval = y_pred.eval()
## LOSS
loss_value_new = loss.eval()
loss_ratio = (loss_value - loss_value_new) / loss_value
loss_value = loss_value_new
## MSE
mse_value_new = mse(ys_eval, y_pred_eval)
mse_ratio = (mse_value - mse_value_new) / mse_value
mse_value = mse_value_new
## AUC
auc_value = auc(ys_eval, y_pred_eval)
if (i+1) % 100 == 0:
self.logger.info("Iterator[%05d]: LOSS[%.6f], MSE[%.6f], AUC[%.6f]" % (i+1, loss_value, mse_value, auc_value))
if loss_ratio <= self._tol or mse_ratio <= self._tol or (i+1) == self._max_iter:
self.logger.info("Iterator[%05d]: LOSS[%.6f], MSE[%.6f], AUC[%.6f]" % (i+1, loss_value, mse_value, auc_value))
break
sess.run(train_step)
## save model
self._w0 = np.matrix(w0.eval())
self._w = np.matrix(w.eval())
self._v = np.matrix(v.eval())
return self
def run():
data_home = TASK_DATA_HOME + "/algorithm_test/kaggle/hr_analytics"
df = pd.read_csv(data_home + "/hr.csv")
df.reset_index(inplace=True)
id = "index"
features_continuous = [
"satisfaction_level",
"last_evaluation",
"number_project",
"average_montly_hours",
"time_spend_company",
]
features_discrete = [
"Work_accident", "promotion_last_5years", "sales", "salary"
]
df["__label__"] = df["left"].apply(lambda x: 1 if int(x) == 1 else 0)
label = "__label__"
df_coding = df[[id, label]]
for feature in features_continuous:
df_coding_f = discretization_coding(df, id, feature)
df_coding = df_coding.merge(df_coding_f, how="inner", on=id)
for feature in features_discrete:
df_coding_f = one_hot_coding(df, id, feature)
df_coding = df_coding.merge(df_coding_f, how="inner", on=id)
x_columns = filter(lambda x: x not in (id, label), df_coding.columns)
y_columns = [label]
df_train, df_test = train_test_split(df_coding, test_size=0.2)
X_train = df_train[x_columns].reset_index(drop=True)
X_test = df_test[x_columns].reset_index(drop=True)
y_train = df_train[[label]].reset_index(drop=True)
y_test = df_test[[label]].reset_index(drop=True)
## compare lr, gbdt, fm and deeplearning
print "############ lr ############"
algo_lr = LogisticRegressionAlgorithm()
mse_lr, auc_lr, threshold_lr = kfold_cross_validation(
algo_lr, X_train, y_train, 5)
algo_lr = algo_lr.fit(X_train, y_train)
y_predict_lr = pd.DataFrame(algo_lr.predict(X_test), columns=["__label__"])
y_predict_lr["__class__"] = y_predict_lr["__label__"].apply(
lambda x: 1 if x >= threshold_lr else 0)
print "lr : mse[%.6f], auc[%.6f]" % (mse(
y_test,
y_predict_lr["__label__"]), auc(y_test, y_predict_lr["__label__"]))
y_columns = filter(lambda x: x not in ("item"), df_item_coding.columns)
df_train, df_test = train_test_split(df_data, test_size=0.2)
label = "Action"
X_train = df_train[x_columns].reset_index(drop=True)
X_test = df_test[x_columns].reset_index(drop=True)
y_train = df_train[[label]].reset_index(drop=True)
y_test = df_test[[label]].reset_index(drop=True)
## compare lr, gbdt, fm and deeplearning
print "############ lr ############"
algo_lr = LogisticRegressionAlgorithm()
algo_lr = algo_lr.fit(X_train, y_train)
y_predict_lr = pd.DataFrame(algo_lr.predict(X_test), columns=["__label__"])
print "lr : mse[%.6f], auc[%.6f]" % (mse(
y_test, y_predict_lr), auc(y_test, y_predict_lr))
print "############ gbdt ############"
algo_gbdt = GradientBoostingRegressor()
algo_gbdt = algo_gbdt.fit(X_train, y_train)
y_predict_gbdt = pd.DataFrame(algo_gbdt.predict(X_test),
columns=["__label__"])
print "gbdt: mse[%.6f], auc[%.6f]" % (mse(
y_test, y_predict_gbdt), auc(y_test, y_predict_gbdt))
print "############ fm ############"
algo_fm = FactorizationMachineAlgorithm()
algo_fm = algo_fm.fit(X_train, y_train)
y_predict_fm = pd.DataFrame(algo_fm.predict(X_test), columns=["__label__"])
print "fm : mse[%.6f], auc[%.6f]" % (mse(
y_test, y_predict_fm), auc(y_test, y_predict_fm))
def run():
data_home = TASK_DATA_HOME + "/algorithm_test/kaggle/creditcardfraud"
df = pd.read_csv(data_home + "/creditcard.sub.csv")
df.reset_index(inplace=True)
id = "index"
features_continuous = filter(lambda x: x not in ("Class", id),
df.columns.tolist())
features_discrete = []
df["__label__"] = df["Class"].apply(lambda x: 1 if int(x) == 1 else 0)
label = "__label__"
df_coding = df[[id, label]]
for feature in features_continuous:
df_coding_f = discretization_coding(df, id, feature)
df_coding = df_coding.merge(df_coding_f, how="inner", on=id)
for feature in features_discrete:
df_coding_f = one_hot_coding(df, id, feature)
df_coding = df_coding.merge(df_coding_f, how="inner", on=id)
df_coding = df_coding.sample(frac=1).reset_index(drop=True)
x_columns = filter(lambda x: x not in (id, label), df_coding.columns)
y_columns = [label]
X = df_coding[x_columns]
y = df_coding[y_columns]
index_pos = y[y["__label__"] == 1].index.tolist()
index_neg = y[y["__label__"] == 0].index.tolist()
index_pos_train, index_pos_test = train_test_split(index_pos,
test_size=0.2)
index_neg_train, index_neg_test = train_test_split(index_neg,
test_size=0.2)
X_train = X.loc[index_pos_train + index_neg_train, :]
X_test = X.loc[index_pos_test + index_neg_test, :]
y_train = y.loc[index_pos_train + index_neg_train, :]
y_test = y.loc[index_pos_test + index_neg_test, :]
# print "############ lr ############"
# algo_lr = LogisticRegressionAlgorithm()
# algo_lr = algo_lr.fit(X_train, y_train)
# algo_lr.predict(X_test)
# y_predict_lr = pd.DataFrame(algo_lr.predict(X_test), columns=["__label__"])
# print "lr : mse[%.6f], auc[%.6f]" % (mse(y_test, y_predict_lr), auc(y_test, y_predict_lr))
#
# print "############ gbdt ############"
# algo_gbdt = GradientBoostingRegressor()
# algo_gbdt = algo_gbdt.fit(X_train, y_train)
# y_predict_gbdt = pd.DataFrame(algo_gbdt.predict(X_test), columns=["__label__"])
# print "gbdt: mse[%.6f], auc[%.6f]" % (mse(y_test, y_predict_gbdt), auc(y_test, y_predict_gbdt))
#
# print "############ fm ############"
# algo_fm = FactorizationMachineAlgorithm()
# algo_fm = algo_fm.fit(X_train, y_train)
# y_predict_fm = pd.DataFrame(algo_fm.predict(X_test), columns=["__label__"])
# print "fm : mse[%.6f], auc[%.6f]" % (mse(y_test, y_predict_fm), auc(y_test, y_predict_fm))
print "############ lp ############"
algo_lp = LittleProbabilityModel(n_jobs=2)
kfold_cross_validation(algo_lp, X_train, y_train, 5)
algo_lp = algo_lp.fit(X_train, y_train)
y_predict_lp = pd.DataFrame(algo_lp.predict(X_test), columns=["__label__"])
print "lp : mse[%.6f], auc[%.6f]" % (mse(
y_test, y_predict_lp), auc(y_test, y_predict_lp))