【競馬AI 12】損失関数を使って穴馬を予測するモデルを構築する

#穴馬の学習

import lightgbm as lgb
import pandas as pd
from sklearn.metrics import roc_curve,roc_auc_score
import numpy as np
from sklearn import metrics
from scipy.misc import derivative

def sigmoid(x):
    return 1 / (1 + np.exp(-x))

def focal_loss(x: np.ndarray, dtrain: lgb.Dataset, gamma: float) -> np.ndarray:
    """損失計算"""
    x = sigmoid(x)
    x[dtrain == 0] = 1 - x[dtrain == 0] 
    return  -1 * (1 - x)**gamma * np.log(x)

def focal_loss_grad_hess(x: np.ndarray, dtrain: lgb.Dataset, gamma: float) -> (np.ndarray, np.ndarray,):
    """Focal Lossのgradientとhessianを返す"""
    t = dtrain.label
    grad = derivative(lambda _x: focal_loss(_x, t, gamma=gamma), x, n=1, dx=1e-6)
    hess = derivative(lambda _x: focal_loss(_x, t, gamma=gamma), x, n=2, dx=1e-6)
    return grad, hess

def split_date(df, test_size):
    sorted_id_list = df.sort_values('race_id').index.unique()
    train_id_list = sorted_id_list[:round(len(sorted_id_list) * (1-test_size))]
    test_id_list = sorted_id_list[round(len(sorted_id_list) * (1-test_size)):]
    train = df.loc[train_id_list]
    test = df.loc[test_id_list]
    return train, test

# モデルファイル
model_file = 'model/model_ana.txt'

# データの読み込み
data = pd.read_csv('encoded/encoded_data.csv')

# ターゲット変数の生成
data['着順'] = ((data['着順'] <= 3) & (data['人気'] >= 6) & (data['オッズ'] >= 20)).astype(int)

# 特徴量とターゲットの分割
train, test = split_date(data, 0.2)
drop_arr=['着順','オッズ','人気','上がり','走破時間','通過順']
X_train = train.drop(drop_arr, axis=1)
# ,'馬の平均着順', '馬の3着内率'
y_train = train['着順']
X_test = test.drop(drop_arr, axis=1)
y_test = test['着順']

# LightGBMデータセットの作成
train_data = lgb.Dataset(X_train, label=y_train)
valid_data = lgb.Dataset(X_test, label=y_test)

weights = len(y_train) / (2 * np.bincount(y_train))
class_weights = {0: weights[0], 1: weights[1]}

gamma = 5.0  # 例としての値
fobj_focal_loss = lambda x,y: focal_loss_grad_hess(x, y, gamma)
params = {
    'metric': 'auc',  # 評価指標
    'learning_rate': 0.05,
    'objective': fobj_focal_loss
    # 他のパラメータ
}

lgb_clf = lgb.train(
    params,
    train_data,
    valid_sets=[valid_data]
)

# 予測の取得
y_pred_train = lgb_clf.predict(X_train)
y_pred = lgb_clf.predict(X_test)

# モデルの評価
print(roc_auc_score(y_test, y_pred))

# モデルの保存
lgb_clf.save_model(model_file)

# 既に訓練したモデル 'model' とテストデータ 'X_test', 'y_test' があると仮定します
# テストデータに対する予測確率を求める
model = lgb.Booster(model_file=model_file)
test_probs = model.predict(X_test)

# AUCを計算する
fpr, tpr, thresholds = metrics.roc_curve(y_test, test_probs, pos_label=1)
test_auc = metrics.auc(fpr, tpr)

print("Test AUC: ", test_auc)