天池二手車價格預測--建模與調參

模型選型

模型調參

1 模型選型

採用了xgboost和lightGBM以及它倆的加權模型

2 模型調參

主要對葉子節點數，學習率以及估計器參數進行調整

    def xgb_model_fit(self,
                      X_train, X_test, y_train, y_test,alg,  useTrainCV=True, cv_folds=5, early_stopping_rounds=50):
        if useTrainCV:
            """訓練集訓練數據"""
            xgb_param = alg.get_xgb_params()
            xgtrain = xgb.DMatrix(X_train, label=y_train)
            cvresult = xgb.cv(xgb_param, xgtrain, num_boost_round=alg.get_params()['n_estimators'], nfold=cv_folds,
                              metrics='mae', early_stopping_rounds=early_stopping_rounds)
            alg.set_params(n_estimators=cvresult.shape[0])

        # 擬合模型
        alg.fit(X_train, y_train, eval_metric='mae')

        # 預測訓練集、測試集
        train_data_df_predictions = alg.predict(X_train)
        test_data_df_predictions = alg.predict(X_test)


        # 迴歸問題評價標--訓練集
        print("training mean_absolute_error is : " )
        print(mean_absolute_error(y_train, train_data_df_predictions))

        #測試集
        print("test mean_absolute_error is : ")
        print(mean_absolute_error(y_test, test_data_df_predictions))

        #特徵重要度
        plt.ylabel('Feature Importance is')
        plot_importance(alg)
        plt.show()
    def bak_log(self,x_list):
        return list(map(lambda x:pow(2.72,x),x_list))

    def light_gbm_model_fit(self,X_train, X_test, y_train, y_test):
        gbm = lgb.LGBMRegressor(objective='regression', num_leaves=175, learning_rate=0.05, n_estimators=20) #num_leaves=31
        gbm.fit(X_train, y_train, eval_set=[(X_test, y_test)], eval_metric='l1', early_stopping_rounds=5)

        print('Start lightgbm  predicting...')
        # 訓練集與測試集預測
        y_train_pred = gbm.predict(X_train, num_iteration=gbm.best_iteration_)

        y_test_pred = gbm.predict(X_test, num_iteration=gbm.best_iteration_)
        # 模型評估
        print('The y_train mae of test prediction is:', mean_absolute_error(DataSearch().bak_log(y_train), DataSearch().bak_log(y_train_pred)))
        print('The y_test mae of test prediction is:', mean_absolute_error(DataSearch().bak_log(y_test), DataSearch().bak_log(y_test_pred)))

        # feature importances
        print('Feature importances:', list(gbm.feature_importances_))


        # 網格搜索，參數優化
        estimator = lgb.LGBMRegressor(num_leaves=175, metrics='mae', max_depth=7, min_child_samples=1000) #num_leaves=64



        param_grid = {
            'learning_rate': [0.01, 0.1, 1],
            'n_estimators': [20, 40]
        }

        gbm_grid = GridSearchCV(estimator, param_grid)

        gbm_grid = gbm_grid.fit(X_train, y_train)
        print("用網格搜索的方式開始進行預測")
        print('Best parameters found by grid search are:', gbm_grid.best_params_)

        # 訓練集與測試集預測
        y_train_pred = gbm_grid.predict(X_train)

        y_test_pred = gbm_grid.predict(X_test)
        # 模型評估
        print('grid search cv  The y_train mae of test prediction is:', mean_absolute_error(y_train, y_train_pred))
        print('grid search cv  The y_test mae of test prediction is:', mean_absolute_error(y_test, y_test_pred))

        # feature importances
        print('Feature importances:', list(gbm_grid.feature_importances_))

結果

The bak log test prediction is: 776.484369

待續