1. 对比标准

1.1 数据集

• 分类：Fashion MNIST（60000条数据784个特征）
• 回归：NYC Taxi fares（60000条数据7个特征）
• 大规模数据集：NYC Taxi fares（2百万条数据7个特征）

PS：本文只进行了分类的对比

1.2 规则

1. 使用基准模型
2. 使用相同参数训练并利用GridSearchCV调参
3. 比较训练和预测耗时、预测分数、可解释性

1.3 版本

xgboost==0.90
lightgbm==2.3.1
catboost==0.21

2. 结果

2.1 准确率

LightGBM>XGBoost>CatBoost

2.2 训练时间和预测时间

CatBoost<LightGBM<XGBoost

2.3 可解释性

XGBoost=LightGBM>Catboost

2.3.1 特征重要性

2.3.2 SHAP值

类别	含义
0	T-shirt/top
1	Trouser
2	Pullover
3	Dress
4	Coat
5	Sandal
6	Shirt
7	Sneaker
8	Bag
9	Ankle boot

XGBoost

LightGBM

CatBoost无法开箱即用

2.3.3 可视化二叉树

XGBoost

LightGBM

CatBoost绘制树函数

3. 总结

比赛选LightGBM，工业选Catboost

4. 代码

https://download.csdn.net/download/lly1122334/12171980

参考文献

1. Battle of the Boosting Algos: LGB, XGB, Catboost
2. Battle of the Boosting Algorithms
3. mlxtend: A library of extension and helper modules for Python’s data analysis and machine learning libraries
4. shap: A game theoretic approach to explain the output of any machine learning model
5. http://www.picnet.com.au/blogs/guido/post/2016/09/22/xgboost-windows-x64-binaries-for-download/
6. Graphviz – Graph Visualization Software Windows Packages

备份

# Feature Engineering
# this cell was adapted from https://www.kaggle.com/mahtieu/nyc-taxi-fare-prediction-data-expl-xgboost
def feature_engineering(df):
    df['pickup_datetime'] = pd.to_datetime(df['pickup_datetime'])
    #Drop rows with null values
    df = df.dropna(how = 'any', axis = 'rows')
    #Free rides, negative fares and passenger count filtering
    df = df[df.eval('(fare_amount > 0) & (passenger_count <= 6)')]
    # Coordinates filtering - Pickup and dropoff locations should be within the limits of NYC
    df = df[(df.pickup_longitude >= -77) &
                  (df.pickup_longitude <= -70) &
                  (df.dropoff_longitude >= -77) &
                  (df.dropoff_longitude <= 70) &
                  (df.pickup_latitude >= 35) &
                  (df.pickup_latitude <= 45) &
                  (df.dropoff_latitude >= 35) &
                  (df.dropoff_latitude <= 45)]

    df.pickup_datetime = df.pickup_datetime.dt.tz_convert('UTC')
    df.pickup_datetime = df.pickup_datetime.dt.tz_convert('America/New_York')

    # Fares may change every year
    df['year'] = df.pickup_datetime.dt.year

    # Different fares during weekdays and weekends
    df['dayofweek'] = df.pickup_datetime.dt.dayofweek

    # Different fares during public holidays
    df['dayofyear'] = df.pickup_datetime.dt.dayofyear

    # Different fares in peak periods and off-peak periods
    df['hourofday'] = df.pickup_datetime.dt.hour

    df = df.drop('pickup_datetime', axis=1)

    # Computes the distance (in miles) between the pickup and the dropoff locations
    df['distance'] = df.apply(
        lambda x: distance.distance((x.pickup_latitude, x.pickup_longitude), (x.dropoff_latitude, x.dropoff_longitude)).miles,
        axis = 1)

    df = df[df.eval('(distance > 0) & (distance < 150)')]
    fare_distance_ratio = (df.fare_amount/df.distance)
    fare_distance_ratio.describe()

    (fare_distance_ratio[fare_distance_ratio < 45]).hist()

    # Drop incoherent fares
    df = df[fare_distance_ratio < 45]
    del fare_distance_ratio

    # Coordinates of the 3 airpots of NYC
    airports = {'jfk': [40.6441666, -73.7822222],
                'laguardia': [40.7747222, -73.8719444],
                'newark': [40.6897222, -74.175]}

    # Computes the distance between the pickup location and the airport
    pickup = df.apply(lambda x: distance.distance((x.pickup_latitude, x.pickup_longitude), (airports.get('jfk'))).miles, axis=1)
    # Computes the distance between the dropoff location and the airport
    dropoff = df.apply(lambda x: distance.distance((x.dropoff_latitude, x.dropoff_longitude), (airports.get('jfk'))).miles, axis=1)
    # Selects the shortest distance
    df['to_jfk'] = pd.concat((pickup, dropoff), axis=1).min(axis=1)

    pickup = df.apply(lambda x: distance.distance((x.pickup_latitude, x.pickup_longitude), (airports.get('laguardia'))).miles, axis=1)
    dropoff = df.apply(lambda x: distance.distance((x.dropoff_latitude, x.dropoff_longitude), (airports.get('laguardia'))).miles, axis=1)
    df['to_laguardia'] = pd.concat((pickup, dropoff), axis=1).min(axis=1)

    pickup = df.apply(lambda x: distance.distance((x.pickup_latitude, x.pickup_longitude), (airports.get('newark'))).miles, axis=1)
    dropoff = df.apply(lambda x: distance.distance((x.dropoff_latitude, x.dropoff_longitude), (airports.get('newark'))).miles, axis=1)
    df['to_newark'] = pd.concat((pickup, dropoff), axis=1).min(axis=1)
    del pickup, dropoff
    return df