使用结构化表格数据对比深度学习和GBDT模型（下）

@timerdeftabNetPretrain(X_train):
"""Pretrain TabNet modelReturn:TabNet pretrainer obj"""tabnet_params=dict(n_d=8, n_a=8, n_steps=3, gamma=1.3,
n_independent=2, n_shared=2,
seed=42, lambda_sparse=1e-3,
optimizer_fn=torch.optim.Adam,
optimizer_params=dict(lr=2e-2,
weight_decay=1e-5                                                  ),
mask_type="entmax",
scheduler_params=dict(max_lr=0.05,
steps_per_epoch=int(X_train.shape[0] /256),
epochs=200,
is_batch_level=True                                                  ),
scheduler_fn=torch.optim.lr_scheduler.OneCycleLR,
verbose=10                        )
pretrainer=TabNetPretrainer(**tabnet_params)
pretrainer.fit(
X_train=X_train.to_numpy(),
eval_set=[X_train.to_numpy()],
max_epochs=100,
patience=10, 
batch_size=256, 
virtual_batch_size=128,
num_workers=1, 
drop_last=True)
returnpretrainer

#MakepredictionsdefmakePredictions(X_test, xgb_model, d1_cnn_model, tabNet_model):
"""Make predictionsReturn:Predictions from each models"""y_xgb_pred=xgb_model.predict(xgb.DMatrix(X_test, feature_names=FEATS))
y_d1_cnn_pred=d1_cnn_model.predict(X_test).reshape(1, -1)[0]
y_tabNet_pred=tabNet_model.predict_proba(X_test.to_numpy())[:,1]
return [y_xgb_pred, y_d1_cnn_pred, y_tabNet_pred]

#Evaluationdefevaluate(y_xgb_pred, y_d1_cnn_pred, y_tabNet_pred) ->None:
"""Evaluate the predictionsProcess:Print ROC AUC and F1 score of each models"""preds= {"XGBoost":y_xgb_pred, "D1 CNN":y_d1_cnn_pred, "TabNet":y_tabNet_pred}
forkeyinpreds:
print("The ROC AUC score of "+str(key) +" model is "+str(round(roc_auc_score(y_test, preds[key]), 4))
             )
forkeyinpreds:
print("The F1 score of "+str(key) +" model at threshold = 0.27 is "+str(round(f1_score(y_test, np.where(preds[key] >0.27, 1, 0)), 4))
             )

#PlotpredictiondistributiondefplotPredictionDistribution(y_xgb_pred, y_d1_cnn_pred, y_tabNet_pred) ->None:
"""Plot histogram of predicted probability distributions of each model"""preds= {"XGBoost":y_xgb_pred, "D1 CNN":y_d1_cnn_pred, "TabNet":y_tabNet_pred}
forkeyinpreds:
plt.hist(preds[key], bins=100)
plt.title(f"Predicted probability distribution of {key}")
plt.show()

ROUNDS=500FEATS= [
"CLAIM3YEARS_Y", "BUS_USE_Y", "AD_BUILDINGS_Y",
"CONTENTS_COVER_Y", "P1_SEX_M", "P1_SEX_N", "BUILDINGS_COVER_Y", 
"P1_POLICY_REFUSED_Y", "APPR_ALARM_Y", "APPR_LOCKS_Y", "FLOODING_Y", 
"NEIGH_WATCH_Y", "SAFE_INSTALLED_Y", "SEC_DISC_REQ_Y", "SUBSIDENCE_Y", 
"LEGAL_ADDON_POST_REN_Y", "HOME_EM_ADDON_PRE_REN_Y", 
"HOME_EM_ADDON_POST_REN_Y", "GARDEN_ADDON_PRE_REN_Y",
"GARDEN_ADDON_POST_REN_Y", "KEYCARE_ADDON_PRE_REN_Y", 
"KEYCARE_ADDON_POST_REN_Y", "HP1_ADDON_PRE_REN_Y", "HP1_ADDON_POST_REN_Y", 
"HP2_ADDON_PRE_REN_Y", "HP2_ADDON_POST_REN_Y", "HP3_ADDON_PRE_REN_Y", 
"HP3_ADDON_POST_REN_Y", "MTA_FLAG_Y", "OCC_STATUS_LP",
"OCC_STATUS_PH", "OCC_STATUS_UN", "OCC_STATUS_WD",
"OWNERSHIP_TYPE_2.0", "OWNERSHIP_TYPE_3.0", "OWNERSHIP_TYPE_6.0", 
"OWNERSHIP_TYPE_7.0", "OWNERSHIP_TYPE_8.0", "OWNERSHIP_TYPE_11.0", 
"OWNERSHIP_TYPE_12.0", "OWNERSHIP_TYPE_13.0", "OWNERSHIP_TYPE_14.0", 
"OWNERSHIP_TYPE_16.0", "OWNERSHIP_TYPE_17.0", 
"OWNERSHIP_TYPE_18.0", "PROP_TYPE_2.0", "PROP_TYPE_3.0", "PROP_TYPE_4.0", 
"PROP_TYPE_7.0", "PROP_TYPE_9.0", "PROP_TYPE_10.0", 
"PROP_TYPE_16.0", "PROP_TYPE_17.0", "PROP_TYPE_18.0", "PROP_TYPE_19.0", 
"PROP_TYPE_20.0", "PROP_TYPE_21.0", "PROP_TYPE_22.0", "PROP_TYPE_23.0", 
"PROP_TYPE_24.0", "PROP_TYPE_25.0", "PROP_TYPE_26.0", "PROP_TYPE_27.0", 
"PROP_TYPE_29.0", "PROP_TYPE_30.0", "PROP_TYPE_31.0", 
"PROP_TYPE_32.0", "PROP_TYPE_37.0", "PROP_TYPE_39.0", 
"PROP_TYPE_40.0", "PROP_TYPE_44.0", "PROP_TYPE_45.0", "PROP_TYPE_47.0", 
"PROP_TYPE_48.0", "PROP_TYPE_51.0", "PROP_TYPE_52.0", "PROP_TYPE_53.0", 
"PAYMENT_METHOD_NonDD", "PAYMENT_METHOD_PureDD", "P1_EMP_STATUS_C", 
"P1_EMP_STATUS_E", "P1_EMP_STATUS_F", "P1_EMP_STATUS_H", "P1_EMP_STATUS_I", 
"P1_EMP_STATUS_N", "P1_EMP_STATUS_R", "P1_EMP_STATUS_S", "P1_EMP_STATUS_U", 
"P1_EMP_STATUS_V", "P1_MAR_STATUS_B", "P1_MAR_STATUS_C", "P1_MAR_STATUS_D", 
"P1_MAR_STATUS_M", "P1_MAR_STATUS_N", "P1_MAR_STATUS_O", "P1_MAR_STATUS_P", 
"P1_MAR_STATUS_S", "P1_MAR_STATUS_W", 
"age", "property_age", "cover_length", "RISK_RATED_AREA_B_imputed", 
"RISK_RATED_AREA_C_imputed", "MTA_FAP_imputed", "MTA_APRP_imputed",
"SUM_INSURED_BUILDINGS", "NCD_GRANTED_YEARS_B", "SUM_INSURED_CONTENTS", 
"NCD_GRANTED_YEARS_C", "SPEC_SUM_INSURED", "SPEC_ITEM_PREM", 
"UNSPEC_HRP_PREM", "BEDROOMS", "MAX_DAYS_UNOCC", "LAST_ANN_PREM_GROSS"        ]
print("Reading the data")
df=pd.read_csv("../input/home-insurance/home_insurance.csv")
print("Preprocessing the data")
X_train, y_train, X_test, y_test=splitData(df, FEATS)
X_train, X_test=standardiseNumericalFeats(X_train, X_test)
print("The ratio of lapse class in training set is "+str(round(y_train.sum()/len(y_train) *100, 2)) +"%"     )
print("The ratio of lapse class in test set is "+str(round(y_test.sum()/len(y_test) *100, 2)) +"%"     )
print("Training XGBoost model")
xgb_model=trainXgbModel(X_train, y_train, X_test, y_test, FEATS, ROUNDS)
print("Training MLP model")
d1_cnn_model=trainD1CnnModel(X_train, y_train)
print("Training TabNet model")
tabNet_model=trainTabNetModel(X_train, y_train, None)
print("Making predictions")
y_xgb_pred, y_d1_cnn_pred, y_tabNet_pred=makePredictions(X_test, xgb_model, d1_cnn_model, tabNet_model)
print("Evaluation of the model")
evaluate(y_xgb_pred, y_d1_cnn_pred, y_tabNet_pred)
print("Prediction distribution")
plotPredictionDistribution(y_xgb_pred, y_d1_cnn_pred, y_tabNet_pred)