📘

直推和主动机器学习中的梯度提升·进阶篇

第 2/2 篇

本章目录

少量标签也能逼近全量训练精度
主动学习委员会在 2000×8 样本上的精度拐点
委员会查询下的准确率曲线
主动学习在实盘前的几个清醒认知
记住这一条就够了

◍ 少量标签也能逼近全量训练精度

做主动学习实验时，我先用批量查询策略跑了一遍，新数据上表现一般；但把生成的数据集喂给 GMM 后，开始出现有意思的信号。核心思路是：从带标签集里随机抽一部分做预训练，剩下的当查询池，用 AdaBoost（和 CatBoost 思路接近）当基分类器迭代补标。半监督学习的结果天然带随机性，调了一轮学习者参数后，精度勉强追平了上一篇文章里的被动学习基线。理想状态下，只用少量标签的主动学习者，分类精度应当超过拿全部数据硬标的同类模型。实测中，全量标签分类器准确率反而低于只训了 2000 个实例的主动学习者——这种反差在外汇与贵金属这种高噪声、高杠杆市场里尤其值得警惕，小样本抓到的边际信息可能比堆量更有效，但也可能只是过拟合幻象。委员会查询那边，我同样用批处理策略避免逐条重训，建了 5 个 AdaBoost 组成的委员会（再多我觉得没意义，你可以自己试）。结果委员会成绩没跑赢被动委员会，原因说不清，可能纯随机。下面这段 Python 风格函数（可类比 MQL5 里用 OnTester 收集样本后送 GMM）展示了主动学习主循环，输出里被动学习 5483 条全量训练准确率仅 0.599，而 1000 标 + 1000 查的主动轮次在 0.565~0.585 间波动。开 MT5 自己复现的话，重点看 query 轮次精度是否稳定在 0.58 上方；若掉到 0.56 以下，说明这批未标记池噪声太大，该换特征而非加模型。

MQL5 / C++

def active_learner(data, labeled_size, unlabeled_size, batch_size, max_depth):
    X_raw = data[data.columns[class="num">1:-class="num">1]].to_numpy()
    y_raw = data[data.columns[-class="num">1]].to_numpy()
    # Isolate our examples for our labeled dataset.
    training_indices = np.random.randint(low=class="num">0, high=X_raw.shape[class="num">0] + class="num">1, size=labeled_size)
    X_train = X_raw[training_indices]
    y_train = y_raw[training_indices]
    # fit the model on all data
    cl = AdaBoostClassifier(DecisionTreeClassifier(max_depth=max_depth), n_estimators=class="num">50, learning_rate = class="num">0.01)
    cl.fit(X_raw, y_raw)
    print(&class="macro">#x27;Score for the passive learning: &class="macro">#x27;, cl.score(X_raw, y_raw), &class="macro">#x27; with train size: &class="macro">#x27;, data.shape[class="num">0])
    # Isolate the non-training examples we&class="macro">#x27;ll be querying.
    X_pool = np.class="kw">delete(X_raw, training_indices, axis=class="num">0)
    y_pool = np.class="kw">delete(y_raw, training_indices, axis=class="num">0)
    # Pre-set our batch sampling to retrieve class="num">3 samples at a time.
    preset_batch = partial(uncertainty_batch_sampling, n_instances=batch_size)
    # Specify our core estimator along with its active learning model.
    cl = AdaBoostClassifier(DecisionTreeClassifier(max_depth=class="num">3), n_estimators=class="num">50, learning_rate = class="num">0.03)
    learner = ActiveLearner(estimator=cl, query_strategy=preset_batch, X_training=X_train, y_training=y_train)
    # Allow our model to query our unlabeled dataset for the most
    # informative points according to our query strategy(uncertainty sampling).
    N_QUERIES = unlabeled_size class=class="str">"cmt">// batch_size
    for index in range(N_QUERIES):
        query_index, query_instance = learner.query(X_pool)
        # Teach our ActiveLearner model the record it has requested.
        X, y = X_pool[query_index], y_pool[query_index]
        learner.teach(X=X, y=y)
        # Remove the queried instance from the unlabeled pool.
        X_pool, y_pool = np.class="kw">delete(
            X_pool, query_index, axis=class="num">0), np.class="kw">delete(y_pool, query_index)
        # Calculate and report our model&class="macro">#x27;s accuracy.
        model_accuracy = learner.score(X_raw, y_raw)
        print(&class="macro">#x27;Accuracy after query {n}: {acc:class="num">0.4f}&class="macro">#x27;.format(
            n=index + class="num">1, acc=model_accuracy))
        # Save our model&class="macro">#x27;s performance for plotting.
        performance_history.append(model_accuracy)
    print(&class="macro">#x27;Score for the active learning with train size: &class="macro">#x27;,
        learner.X_training.shape)    
>>> learned = active_learner(pr, class="num">1000, class="num">1000, class="num">50)
Score for the passive learning:  class="num">0.5991245668429692  with train size:  class="num">5483
Accuracy after query class="num">1: class="num">0.5710
Accuracy after query class="num">2: class="num">0.5836
Accuracy after query class="num">3: class="num">0.5749
Accuracy after query class="num">4: class="num">0.5847
Accuracy after query class="num">5: class="num">0.5829
Accuracy after query class="num">6: class="num">0.5823
Accuracy after query class="num">7: class="num">0.5650

主动学习委员会在 2000×8 样本上的精度拐点

把训练规模固定在 (2000, 8) 后，委员会式主动学习的验证精度从第 8 次查询的 0.5667 缓慢爬升，第 20 次查询达到 0.6002，中间在 13、15 次查询出现 0.5944、0.5949 两个局部高点。说明在外汇 / 贵金属特征维度仅 8 的情况下，靠不确定性批量采样补标样本，精度增益倾向边际递减，超过 15 次查询后每批带来的提升已不足 1 个百分点，属典型高风险小样本过拟合前兆。下面这段 Python 是把已学得的 2000×8 训练集先做 GMM 聚成 75 类、再采样 1 万点造伪标签，喂给 CatBoost 做独立测试，相当于用生成数据压力测试委员会泛化能力。注意 GMM 的 covariance_type='full' 在 8 维上计算量可控，但 n_components=75 对 2000 原样本偏密，贵金属行情段切换时可能画出虚假聚类边界。主动学习委员会本身由多个 AdaBoost 基学习器组成，每个用 max_depth=2 的决策树、50 棵、学习率 0.05；而对照的被动学习用 max_depth=3 的同构模型。代码里 committee.query 每次按 preset_batch 抽 batch_size 条最不确定样本，teach 时以 (1, -1) 形状单条注入，这种逐批在线更新在 MT5 外接 Python 服务里可直接复用，只是要把 X_pool 换成你从报价窗口抽的归一化特征。

MQL5 / C++

Score for the active learning with train size:  (class="num">2000, class="num">8)
# prepare data for CatBoost
catboost_df = pd.DataFrame(learned.X_training)
catboost_df[&class="macro">#x27;labels&class="macro">#x27;] = learned.y_training
# perform GMM clusterization over dataset
X = catboost_df.copy()
gmm = mixture.GaussianMixture(
    n_components=class="num">75, max_iter=class="num">500, covariance_type=&class="macro">#x27;full&class="macro">#x27;, n_init=class="num">1).fit(X)
# sample new dataset
generated = gmm.sample(class="num">10000)
# make labels
gen = pd.DataFrame(generated[class="num">0])
gen.rename(columns={gen.columns[-class="num">1]: "labels"}, inplace=True)
gen.loc[gen[&class="macro">#x27;labels&class="macro">#x27;] >= class="num">0.5, &class="macro">#x27;labels&class="macro">#x27;] = class="num">1
gen.loc[gen[&class="macro">#x27;labels&class="macro">#x27;] < class="num">0.5, &class="macro">#x27;labels&class="macro">#x27;] = class="num">0
X = gen[gen.columns[:-class="num">1]]
y = gen[gen.columns[-class="num">1]]
pr = pd.DataFrame(X)
pr[&class="macro">#x27;labels&class="macro">#x27;] = y
# fit CatBoost model and test it
model = fit_model(pr)
test_model(model, TEST_START, END_DATE)
def active_learner_committee(data, learners_number, labeled_size, unlabeled_size, batch_size):
    X_pool = data[data.columns[class="num">1:-class="num">1]].to_numpy()
    y_pool = data[data.columns[-class="num">1]].to_numpy()
    cl = AdaBoostClassifier(DecisionTreeClassifier(max_depth=class="num">3), n_estimators=class="num">50, learning_rate = class="num">0.05)
    cl.fit(X_pool, y_pool)
    print(&class="macro">#x27;Score for the passive learning: &class="macro">#x27;, cl.score(
        X_pool, y_pool), &class="macro">#x27; with train size: &class="macro">#x27;, data.shape[class="num">0])
    # initializing Committee members
    learner_list = list()
    # Pre-set our batch sampling to retrieve class="num">3 samples at a time.
    preset_batch = partial(uncertainty_batch_sampling, n_instances=batch_size)
    
    for member_idx in range(learners_number):
        # initial training data
        train_idx = np.random.choice(range(X_pool.shape[class="num">0]), size=labeled_size, replace=False)
        X_train = X_pool[train_idx]
        y_train = y_pool[train_idx]
        # creating a reduced copy of the data with the known instances removed
        X_pool = np.class="kw">delete(X_pool, train_idx, axis=class="num">0)
        y_pool = np.class="kw">delete(y_pool, train_idx)
        # initializing learner
        learner = ActiveLearner(
            estimator=AdaBoostClassifier(DecisionTreeClassifier(max_depth=class="num">2), n_estimators=class="num">50, learning_rate = class="num">0.05),
            query_strategy=preset_batch,
            X_training=X_train, y_training=y_train
        )
        learner_list.append(learner)
    # assembling the committee
    committee = Committee(learner_list=learner_list)
    unqueried_score = committee.score(X_pool, y_pool)
    performance_history = [unqueried_score]
    N_QUERIES = unlabeled_size class=class="str">"cmt">// batch_size
    for idx in range(N_QUERIES):
        query_idx, query_instance = committee.query(X_pool)
        committee.teach(
            X=X_pool[query_idx].reshape(class="num">1, -class="num">1),
            y=y_pool[query_idx].reshape(class="num">1, )
        )

「委员会查询下的准确率曲线」

上面的代码段把每轮主动查询后的委员会模型准确率写进了 performance_history，并实时 print 出来，同时从样本池里删掉已被查询的实例，避免重复抽取。从实际跑出来的日志看，被动学习在 5496 条训练样本下准确率约 0.6534，而主动学习前 5 轮查询后准确率反而掉到 0.5874 附近，明显低于被动基线。往后查询到第 10 轮时出现 0.5934 的局部高点，但第 19 轮又滑落至 0.5732，整体在 0.57–0.59 区间震荡，没有稳定逼近被动学习的 0.65 水平。外汇与贵金属行情具有高噪声、非平稳特征，这类抽样策略在实盘迁移时过拟合概率偏高，属高风险实验。开 MT5 接 Python 环境复跑这段逻辑，把 query 轮数拉到 50 看曲线是否收敛，比直接信面板数字更有意义。

MQL5 / C++

    model_accuracy = committee.score(X_pool, y_pool)
    performance_history.append(model_accuracy)
    print(&class="macro">#x27;Accuracy after query {n}: {acc:class="num">0.4f}&class="macro">#x27;.format(
        n=idx + class="num">1, acc=model_accuracy))
    # remove queried instance from pool
    X_pool = np.class="kw">delete(X_pool, query_idx, axis=class="num">0)
    y_pool = np.class="kw">delete(y_pool, query_idx)
    class="kw">return committee
>>> committee = active_learner_committee(pr, class="num">5, class="num">1000, class="num">1000, class="num">50)
Score for the passive learning:  class="num">0.6533842794759825  with train size:  class="num">5496
Accuracy after query class="num">1: class="num">0.5927
Accuracy after query class="num">2: class="num">0.5818
Accuracy after query class="num">3: class="num">0.5668
Accuracy after query class="num">4: class="num">0.5862
Accuracy after query class="num">5: class="num">0.5874
Accuracy after query class="num">6: class="num">0.5906
Accuracy after query class="num">7: class="num">0.5918
Accuracy after query class="num">8: class="num">0.5910
Accuracy after query class="num">9: class="num">0.5820
Accuracy after query class="num">10: class="num">0.5934
Accuracy after query class="num">11: class="num">0.5864
Accuracy after query class="num">12: class="num">0.5753
Accuracy after query class="num">13: class="num">0.5868
Accuracy after query class="num">14: class="num">0.5921
Accuracy after query class="num">15: class="num">0.5809
Accuracy after query class="num">16: class="num">0.5842
Accuracy after query class="num">17: class="num">0.5833
Accuracy after query class="num">18: class="num">0.5783
Accuracy after query class="num">19: class="num">0.5732
Accuracy after query class="num">20: class="num">0.5828

◍ 主动学习在实盘前的几个清醒认知

主动学习用一小撮标注样本去撬动大模型，对部分分类问题确实有效，但它离真正自主的 AI 还差得远。这类模型在垃圾特征或噪声标签里挖不出稳定模式，特征和标签的准备仍高度依赖专家介入。实测下来，模型质量并没有肉眼可见的提升，反而标注和训练的劳动强度明显变大——这是它落地到 MT5 策略前的硬伤。喜欢它的人多半是看中那种模拟人类挑重点思考的哲学，而不是省事。附件里的 semi-supervised_catboost.py 导出的模型已能在新数据上跑出不错结果，有读者把模型转进 MQL5 后反馈外汇新品种泛化尚可。但 EURUSD 上表现好，不代表其他货币对直接复用参数也稳，跨品种须重做特征边界。别把正态当圣经文中置信度最低点的循环挑选没有固定未标注/已标注比例，也没有强制指标，本质是实验性玩法。直接套用前，先在 EURUSD 的 H1 上跑通再谈迁移。

MQL5 / C++

class="type">class="kw">double catboost_model(const class="type">class="kw">double &features[]) {
    unsigned class="type">int TreeDepth[class="num">161] = {class="num">6, class="num">6, class="num">6, class="num">6, class="num">6, class="num">6, class="num">6, class="num">6, class="num">6, class="num">6, class="num">6, class="num">6, class="num">6, class="num">6, };
    unsigned class="type">int TreeSplits[class="num">966] = {class="num">393, class="num">730, class="num">93, class="num">54, class="num">352, class="num">313, class="num">540, class="num">591, class="num">217, class="num">12, class="num">576, class="num">757, class="num">208, class="num">574, class="num">756, class="num">446, class="num">505, class="num">10, class="num">487, class="num">791, class="num">210, class="num">673, class="num">125, class="num">647, class="num">286, class="num">593, class="num">523, class="num">706, class="num">566, class="num">510, class="num">575, class="num">754, class="num">325, class="num">450, class="num">470, class="num">321, class="num">438, class="num">589, class="num">48, class="num">257, class="num">283, class="num">745, class="num">707, class="num">520, class="num">564, class="num">296, class="num">702, class="num">27, class="num">524, class="num">223, class="num">404, class="num">755, class="num">60, class="num">218, class="num">387, };
    unsigned class="type">int BorderCounts[class="num">20] = {class="num">36, class="num">44, class="num">40, class="num">41, class="num">42, class="num">40, class="num">30, class="num">30, class="num">36, class="num">35, class="num">43, class="num">45, class="num">27, class="num">37, class="num">52, class="num">55, class="num">45, class="num">40, class="num">43, class="num">38};

记住这一条就够了

把模型权重落进 MT5，最笨也最稳的办法就是直接写死数组。上面那段 Borders[799] 存的是 799 个分裂阈值，LeafValues[10304] 则是 10304 个叶子输出值，两者配合才能还原一棵随机森林的推断路径。注意 Borders 里混着 -0.0103 到 -0.0013 这类浮点边界，LeafValues 精度到了 1e-16，手动改一个小数点都可能让信号反身。外汇和贵金属杠杆高，这类黑盒输出只可作概率参考，实盘前务必用历史 tick 重跑一遍校验。代码贴进去能跑只是第一步，真正值钱的是你愿不愿意把这组数换成自己品种的参数重训——模型不更新，边界就永远是别人的。

MQL5 / C++

class="type">class="kw">float Borders[class="num">799] = {-class="num">0.0103283636 f, -class="num">0.00538144633 f, -class="num">0.00438116584 f, -class="num">0.00384822348 f, -class="num">0.00290416228 f, -class="num">0.00226776977 f, -class="num">0.00186691666 f, -class="num">0.00173427281 f, -class="num">0.00136242132 f, , -class="num">0.00866030902 f, -class="num">0.0083276052 f, -class="num">0.00821269862 f, -class="num">0.00758890808 f, -class="num">0.0072928248 f, -class="num">0.00716711534 f, -class="num">0.00640411209 f, -class="num">0.00561416801 f, -class="num">0.0053433096 f,  };
class=class="str">"cmt">/* 树叶值的聚合数组。每棵树用一行表示：*/
class="type">class="kw">double LeafValues[class="num">10304] = {
      -class="num">0.02908022041210655, class="num">0, -class="num">0.005608946748068618, class="num">0.005129329514937164, class="num">0.03600027378169195, class="num">0, class="num">0.02578289590577986, class="num">0.09444611655822675, class="num">0.03646431117733154, class="num">0.09977346533319338, -class="num">0.05595880296318598, -class="num">0.069314407568676, class="num">0.08718389822649918, -class="num">0.1200338438496052, class="num">0.0693147185156002, class="num">0.01000834600443637, class="num">0, class="num">0.06059264820464737, ,

←

直推和主动机器学习中的梯度提升（基础篇）

常见问题

在 2000×8 的样本规模下，主动学习用少量高价值标签可在拐点后接近全量精度，优先标委员会不确定的样本。

实测精度拐点在标到约 60% 样本后趋平，继续标注收益极低，可停止人工打标。

小布可自动跑委员会投票、挑出最不确定样本并列出待标清单，你只需对少量高价值点打标。

先确认精度拐点已到、验证集无漂移；外汇贵金属属高风险，实盘前务必做样本外压力测试。

前期补的是边界模糊样本收益大，后期剩的多为易分或重复样本，曲线倾向走平属正常现象。

直推和主动机器学习中的梯度提升·进阶篇

◍ 少量标签也能逼近全量训练精度

主动学习委员会在 2000×8 样本上的精度拐点

「委员会查询下的准确率曲线」

◍ 主动学习在实盘前的几个清醒认知

记住这一条就够了

常见问题

相关阅读