kaggle学习笔记-otto-baseline5-LGBM的使用

数据处理

import polars as pl

train = pl.read_parquet('../input/otto-train-and-test-data-for-local-validation/test.parquet')
train_labels = pl.read_parquet('../input/otto-train-and-test-data-for-local-validation/test_labels.parquet')

def add_action_num_reverse_chrono(df):
    return df.select([
        pl.col('*'),
        pl.col('session').cumcount().reverse().over('session').alias('action_num_reverse_chrono')
    ])

def add_session_length(df):
    return df.select([
        pl.col('*'),
        pl.col('session').count().over('session').alias('session_length')
    ])

def add_log_recency_score(df):
    linear_interpolation = 0.1 + ((1-0.1) / (df['session_length']-1)) * (df['session_length']-df['action_num_reverse_chrono']-1)
    return df.with_columns(pl.Series(2**linear_interpolation - 1).alias('log_recency_score')).fill_nan(1)

def add_type_weighted_log_recency_score(df):
    type_weights = {0:1, 1:6, 2:3}
    type_weighted_log_recency_score = pl.Series(df['log_recency_score'] / df['type'].apply(lambda x: type_weights[x]))
    return df.with_column(type_weighted_log_recency_score.alias('type_weighted_log_recency_score'))

def apply(df, pipeline):
    for f in pipeline:
        df = f(df)
    return df


pipeline = [add_action_num_reverse_chrono, add_session_length, add_log_recency_score, add_type_weighted_log_recency_score]

train = apply(train, pipeline)


train.head()看看数据先

现在我们需要稍微处理一下我们的标签，并将它们合并到我们的训练集中。

type2id = {"clicks": 0, "carts": 1, "orders": 2}
train_labels = train_labels.explode('ground_truth').with_columns([
    pl.col('ground_truth').alias('aid'),
    pl.col('type').apply(lambda x: type2id[x])
])[['session', 'type', 'aid']]

train_labels = train_labels.with_columns([
    pl.col('session').cast(pl.datatypes.Int32),
    pl.col('type').cast(pl.datatypes.UInt8),
    pl.col('aid').cast(pl.datatypes.Int32)
])

train_labels = train_labels.with_column(pl.lit(1).alias('gt'))
train = train.join(train_labels, how='left', on=['session', 'type', 'aid']).with_column(pl.col('gt').fill_null(0))

train.head()

我们现在有了预处理的数据集，一个具有基本信息的列，这意味着我们的 Ranker 唯一缺少的是…如何将各个行分组到会话中的信息

def get_session_lenghts(df):
    return df.groupby('session').agg([
        pl.col('session').count().alias('session_length')
    ])['session_length'].to_numpy()

session_lengths_train = get_session_lenghts(train)

模型训练

from lightgbm.sklearn import LGBMRanker
ranker = LGBMRanker(
    objective="lambdarank",
    metric="ndcg",
    boosting_type="dart",
    n_estimators=20,
    importance_type='gain',
)


train.columns

feature_cols = ['aid', 'type', 'action_num_reverse_chrono', 'session_length', 'log_recency_score', 'type_weighted_log_recency_score']
target = 'gt'



ranker = ranker.fit(
    train[feature_cols].to_pandas(),
    train[target].to_pandas(),
    group=session_lengths_train,
)

在测试数据上进行预测

test = pl.read_parquet('../input/otto-full-optimized-memory-footprint/test.parquet')
test = apply(test, pipeline)
scores = ranker.predict(test[feature_cols].to_pandas())

创建提交版本的CSV文件

test = test.with_columns(pl.Series(name='score', values=scores))
test_predictions = test.sort(['session', 'score'], reverse=True).groupby('session').agg([
    pl.col('aid').limit(20).list()
])


session_types = []
labels = []

for session, preds in zip(test_predictions['session'].to_numpy(), test_predictions['aid'].to_numpy()):
    l = ' '.join(str(p) for p in preds)
    for session_type in ['clicks', 'carts', 'orders']:
        labels.append(l)
        session_types.append(f'{session}_{session_type}')
submission = pl.DataFrame({'session_type': session_types, 'labels': labels})
submission.write_csv('submission.csv')