DataWhale AI夏令营 从零入门 AI for Science（AI+药物）针对lightgbm的改进

低Remaining范围样本高权重

• 通过给低Remaining范围样本更高的权重，模型会更加关注这些样本，从而提高对这些样本的预测准确性。这对于处理不平衡数据集特别有用，因为模型通常会倾向于多数类样本，而忽略少数类样本。
• 在某些情况下，低Remaining范围样本可能包含重要的信息或特征。通过增加这些样本的权重，模型可以更好地学习这些特征，从而减少偏差。
• 通过关注低Remaining范围样本，模型可以更好地适应不同的数据分布，从而提高泛化能力。这意味着模型在处理新数据时表现会更好。
• 在训练过程中，损失函数会根据样本权重进行调整。高权重的低Remaining范围样本会对损失函数产生更大的影响，从而引导模型更好地拟合这些样本。

weight_ls = np.array(feats['mRNA_remaining_pct'].apply(lambda x:2 if ((x<=30)and(x>=0)) else 1))

使用官方评价指标作为损失函数

替换root_mean_squared_error

# calculate_metrics函数用于计算评估指标
def calculate_metrics(preds, data, threshold=30):
    y_pred = preds
    y_true = data.get_label()
    mae = np.mean(np.abs(y_true - y_pred))
    # if mae < 0: mae = 0
    # elif mae >100: mae = 100

    y_true_binary = ((y_true <= threshold) & (y_true >= 0)).astype(int)
    y_pred_binary = ((y_pred <= threshold) & (y_pred >= 0)).astype(int)

    mask = (y_pred >= 0) & (y_pred <= threshold)
    range_mae = (
        mean_absolute_error(y_true[mask], y_pred[mask]) if np.sum(mask) > 0 else 100
    )
    # if range_mae < 0: range_mae = 0
    # elif range_mae >100: range_mae = 100

    # precision = precision_score(y_true_binary, y_pred_binary, average="binary")
    # recall = recall_score(y_true_binary, y_pred_binary, average="binary")

    if np.sum(y_pred_binary) > 0:
        precision = (np.array(y_pred_binary) & y_true_binary).sum()/np.sum(y_pred_binary)
    else:
        precision = 0
    if np.sum(y_true_binary) > 0:
        recall = (np.array(y_pred_binary) & y_true_binary).sum()/np.sum(y_true_binary)
    else:
        recall = 0

    if precision + recall == 0:
        f1 = 0
    else:
        f1 = 2 * precision * recall / (precision + recall)
    score = (1 - mae / 100) * 0.5 + (1 - range_mae / 100) * f1 * 0.5
    return "custom_score", score, True  # True表示分数越高越好

自适应学习率

自适应学习率可以根据训练过程中的反馈动态调整学习率，使模型在训练初期快速收敛，同时在训练后期避免震荡，从而提高训练效率。

# adaptive_learning_rate函数用于自适应学习率
def adaptive_learning_rate(decay_rate=0.8, patience=50):
    best_score = float("-inf")  # 初始化为负无穷,因为分数越高越好
    wait = 0

    def callback(env):
        nonlocal best_score, wait
        current_score = env.evaluation_result_list[-1][2]  # 假设使用的是最后一个评估指标
        current_lr =  env.model.params.get('learning_rate')

        if current_score > best_score: 
            best_score = current_score
            # wait = 0 # 需要连续的score没有上升
        else:
            wait += 1

        if wait >= patience:
            new_lr = float(current_lr) * decay_rate
            wait = 0
            env.model.params['learning_rate'] = new_lr
            print(f"Learning rate adjusted to {env.model.params.get('learning_rate')}")

    return callback

多折交叉训练

将数据集分成多个折叠（folds），每个折叠都作为一次验证集，其余的作为训练集，可以有效减少模型的过拟合风险。模型在不同的训练集和验证集上进行训练和评估，有助于提高模型的泛化能力。

# train函数用于训练模型
def train(feats, n_original):
    # 定义k折交叉验证
    n_splits = 10
    kf = KFold(n_splits=n_splits, shuffle=True, random_state=42)
    # 开始k折交叉验证
    gbms = []
    for fold, (train_idx, val_idx) in enumerate(
        kf.split(feats.iloc[:n_original, :]), 1
    ):
        # 准备训练集和验证集
        X_train, X_val = feats.iloc[train_idx, :-1], feats.iloc[val_idx, :-1]
        y_train, y_val = feats.iloc[train_idx, -1], feats.iloc[val_idx, -1]
        w_train = weight_ls[train_idx]
        

        # 创建LightGBM数据集
        train_data = lgb.Dataset(X_train, label=y_train, weight=w_train)
        val_data = lgb.Dataset(X_val, label=y_val, reference=train_data)

        boost_round = 25000
        early_stop_rounds = int(boost_round*0.1)

        # 显示metric
        lgb_log = lgb.log_evaluation(period=200, show_stdv=True)
        lgb_stop = lgb.early_stopping(stopping_rounds=early_stop_rounds, first_metric_only=True, verbose=True, min_delta=0.00001)

        # 设置LightGBM参数
        params = {
            "boosting_type": "gbdt",
            "objective": "regression",
            "metric": "None",
            # "metric": "root_mean_squared_error",
            "max_depth": 8,
            "num_leaves": 63,
            "min_data_in_leaf": 2,
            "learning_rate": 0.05,
            "feature_fraction": 0.9,
            "lambda_l1": 0.1,
            "lambda_l2": 0.2,
            "verbose": -1, # -1时不输出
            "early_stopping_round": early_stop_rounds,
            "num_threads": 8,
        }

        # 在训练时使用自适应学习率回调函数
        adaptive_lr = adaptive_learning_rate(decay_rate=0.9, patience=1000)
        gbm = lgb.train(
            params,
            train_data,
            num_boost_round=boost_round,
            valid_sets=[val_data],
            feval=calculate_metrics,  # 将自定义指标函数作为feval参数传入
            # callbacks=[print_validation_result, adaptive_lr, lgb_log, lgb_stop],
            callbacks=[adaptive_lr, lgb_log, lgb_stop],
        )
        valid_score = gbm.best_score["valid_0"]["custom_score"]
        print(f"best_valid_score: {valid_score}")
        gbms.append(gbm)

    return gbms

优化结果

比赛得分获得0.03的提升