PyCaretでJリーグ来場者予測

df_j_league = None
for i in range(2010, 2023):
    url = f'https://data.j-league.or.jp/SFMS01/search?competition_years={i}&tv_relay_station_name='
    dfs = pd.read_html(url)
    df_j_league = pd.concat([df_j_league, dfs[0]], axis=0)
    time.sleep(5)

def encode(_df, col):
    _df[col + '_cos'] = np.cos(2 * np.pi * _df[col] / _df[col].max())
    _df[col + '_sin'] = np.sin(2 * np.pi * _df[col] / _df[col].max())
    return _df

def make_train_date(df):
    tmp = df
    tmp['datetime'] = tmp['年度'].astype(str) + '/' + tmp['試合日'].astype(str).str.split('(', expand=True)[0] + ' ' + tmp['K/O時刻']
    tmp['datetime'] = pd.to_datetime(tmp['datetime'], format='%Y/%m/%d %H:%M')
    tmp['year'] = tmp['datetime'].dt.year
    tmp['month'] = tmp['datetime'].dt.month
    tmp['day'] = tmp['datetime'].dt.day
    tmp['dow'] = tmp['datetime'].dt.dayofweek
    tmp['hour'] = tmp['datetime'].dt.hour + tmp['datetime'].dt.minute/60
    tmp['covid'] = 0
    tmp.loc[tmp['year']>=2020, 'covid'] = 1

    tmp = encode(tmp, 'dow')
    tmp = encode(tmp, 'hour')
    tmp = encode(tmp, 'day')
    tmp = encode(tmp, 'month')
    return tmp

# 特徴量エンジニアリング
df = make_train_date(df_j_league)

# 訓練データ
train = df[~(df['入場者数'].isna())].copy()
# 2022年試合予測のためのデータ
test = df[(df['入場者数'].isna())&(df['year']==2022)].copy()

feature = ['year', 'month_cos', 'month_sin', 'day_cos', 'day_sin', 'dow_cos', 'dow_sin', 
           'hour_cos', 'hour_sin', '大会', 'ホーム', 'アウェイ', 'スタジアム', 'covid']
target = '入場者数'
exp_name = setup(data=train[feature+[target]], target=target, 
                 normalize=True, 
                 numeric_features=['year', 'month_cos', 'month_sin', 'day_cos', 'day_sin', 
                                   'dow_cos', 'dow_sin', 'hour_cos', 'hour_sin'])

best_model = compare_models()

rf = create_model('rf')
tuned_rf = tune_model(rf)
predictions = predict_model(tuned_rf , data=test)