{"id":841,"date":"2024-07-15T09:17:21","date_gmt":"2024-07-15T01:17:21","guid":{"rendered":"http:\/\/123.60.176.65\/?p=841"},"modified":"2024-07-15T09:17:21","modified_gmt":"2024-07-15T01:17:21","slug":"%e8%8e%b7%e5%be%97%e7%ae%80%e8%a1%97%e5%b8%82%e5%9c%ba%e9%a2%84%e6%b5%8b%e5%a4%a7%e8%b5%9b%e9%87%91%e7%89%8c%e7%9a%84%e9%a2%84%e6%b5%8b%e7%ad%96%e7%95%a5%e6%a8%a1%e5%9e%8b","status":"publish","type":"post","link":"https:\/\/laoyulaoyu.com\/index.php\/2024\/07\/15\/%e8%8e%b7%e5%be%97%e7%ae%80%e8%a1%97%e5%b8%82%e5%9c%ba%e9%a2%84%e6%b5%8b%e5%a4%a7%e8%b5%9b%e9%87%91%e7%89%8c%e7%9a%84%e9%a2%84%e6%b5%8b%e7%ad%96%e7%95%a5%e6%a8%a1%e5%9e%8b\/","title":{"rendered":"\u83b7\u5f97\u7b80\u8857\u5e02\u573a\u9884\u6d4b\u5927\u8d5b\u91d1\u724c\u7684\u9884\u6d4b\u7b56\u7565\u6a21\u578b"},"content":{"rendered":"\n
\"\"<\/figure>\n\n\n\n
\n
\u5199\u5728\u524d\u9762\u7684\u8bdd\uff1a<\/strong>\u201c\u4f4e\u4e70\u9ad8\u5356\u201d\u542c\u8d77\u6765\u5f88\u7b80\u5355\u2026\u2026<\/mark>\u4e8b\u5b9e\u4e0a\uff0c\u4ea4\u6613\u83b7\u5229\u4e00\u76f4\u662f\u4e00\u4e2a\u96be\u4ee5\u89e3\u51b3\u7684\u95ee\u9898\uff0c\u5728\u5f53\u4eca\u77ac\u606f\u4e07\u53d8\u4e14\u590d\u6742\u7684\u91d1\u878d\u5e02\u573a\u4e2d\u66f4\u662f\u5982\u6b64\u3002\u7535\u5b50\u4ea4\u6613\u5141\u8bb8\u5728\u51e0\u5206\u4e4b\u4e00\u79d2\u5185\u53d1\u751f\u6570\u5343\u7b14\u4ea4\u6613\uff0c\u4ece\u800c\u4e3a\u5b9e\u65f6\u5bfb\u627e\u548c\u5229\u7528\u4ef7\u683c\u5dee\u5f02\u63d0\u4f9b\u4e86\u51e0\u4e4e\u65e0\u9650\u7684\u673a\u4f1a\u3002\u4eca\u5929\u4e3a\u5927\u5bb6\u4ecb\u7ecd\u4e00\u5219\u6709\u5173\u7b80\u8857\u5e02\u573a\u9884\u6d4b\u5927\u8d5b<\/mark>\uff08Jane Street Market Prediction\uff09\u91cc\u83b7\u5f97\u91d1\u724c\u7684\u65b9\u6848<\/mark>\u3002<\/pre>\n<\/blockquote>\n\n\n\n
\u4e00\u3001\u6bd4\u8d5b\u80cc\u666f<\/h3>\n\n\n\n
\u6bd4\u8d5b\u7684\u4e3b\u529e\u65b9\u4e3a\u7b80\u8857\u96c6\u56e2\uff0c\u6b64\u6b21\u8d5b\u4e8b\u5171\u6709 29182 \u540d\u53c2\u8d5b\u8005\uff0c4245 \u652f\u961f\u4f0d\u53c2\u4e0e\uff0c\u6536\u5230\u4e86 7211 \u4efd\u6295\u7a3f\u3002\u53c2\u8d5b\u8005\u9700\u5728\u672c\u6b21\u6311\u6218\u8d5b\u7684\u524d\u4e09\u4e2a\u6708\uff0c\u5229\u7528\u6765\u81ea\u5168\u7403\u4e3b\u8981\u8bc1\u5238\u4ea4\u6613\u6240\u7684\u5e02\u573a\u6570\u636e\u6784\u5efa\u81ea\u8eab\u7684\u91cf\u5316\u4ea4\u6613\u6a21\u578b\uff0c\u4ee5\u8fbe\u6210\u6536\u76ca\u6700\u5927\u5316\u3002\u7d27\u63a5\u7740\uff0c\u4e3b\u529e\u65b9\u4f1a\u6d4b\u9a8c\u6a21\u578b\u5bf9\u672a\u6765\u5e02\u573a\u6536\u76ca\u7684\u9884\u6d4b\u80fd\u529b\uff0c\u5e76\u5bf9\u6a21\u578b\u80fd\u529b\u4e88\u4ee5\u6392\u884c\u3002<\/p>\n\n\n\n
\u901a\u5e38\u6765\u8bb2\uff0c\u82e5\u80fd\u751f\u6210\u4e00\u4e2a\u5177\u5907\u9ad8\u5ea6\u9884\u6d4b\u6027\u7684\u6a21\u578b\u6765\u9009\u5b9a\u6b63\u786e\u7684\u4ea4\u6613\uff0c\u90a3\u5b83\u4eec\u4e5f\u4f1a\u5728\u53d1\u9001\u5e02\u573a\u4fe1\u53f7\u65b9\u9762\u5177\u6709\u91cd\u8981\u4f5c\u7528\uff0c\u4fc3\u4f7f\u4ef7\u683c\u66f4\u4e3a\u8d8b\u8fd1\u201c\u516c\u5e73\u201d\u4ef7\u503c\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u66f4\u4f18\u7684\u6a21\u578b\u610f\u5473\u7740\u5e02\u573a\u5728\u672a\u6765\u4f1a\u66f4\u52a0\u9ad8\u6548\u3002\u7136\u800c\uff0c\u5f00\u53d1\u51fa\u4f18\u8d28\u7684\u6a21\u578b\u5c06\u4f1a\u5145\u6ee1\u6311\u6218\uff0c\u5176\u539f\u56e0\u4f17\u591a\uff0c\u6db5\u76d6\u6781\u4f4e\u7684\u4fe1\u566a\u6bd4\u3001\u6f5c\u5728\u7684\u5197\u4f59\u3001\u5f3a\u70c8\u7684\u7279\u5f81\u76f8\u5173\u6027\u4ee5\u53ca\u96be\u4ee5\u63d0\u51fa\u9002\u5b9c\u7684\u6570\u5b66\u516c\u5f0f\u3002<\/p>\n\n\n\n
\u7b80\u8857\u96c6\u56e2\u8017\u8d39\u6570\u5341\u5e74\u53bb\u5f00\u53d1\u81ea\u8eab\u7684\u4ea4\u6613\u6a21\u578b\u4ee5\u53ca\u673a\u5668\u5b66\u4e60\u89e3\u51b3\u65b9\u6848\uff0c\u7528\u4e8e\u8bc6\u522b\u76c8\u5229\u673a\u9047\u5e76\u8fc5\u901f\u51b3\u5b9a\u662f\u5426\u6267\u884c\u4ea4\u6613\u3002\u8fd9\u4e9b\u6a21\u578b\u52a9\u529b Jane Street \u6bcf\u65e5\u5728\u5168\u7403 200 \u4e2a\u4ea4\u6613\u573a\u6240\u4ea4\u6613\u6570\u5343\u79cd\u91d1\u878d\u4ea7\u54c1\u3002<\/p>\n\n\n\n
\u6bcb\u5eb8\u7f6e\u7591\uff0c\u6b64\u9879\u6311\u6218\u8d5b\u6781\u5927\u5730\u7b80\u5316\u4e86\u7b80\u8857\u96c6\u56e2\u6bcf\u65e5\u5904\u7406\u7684\u91cf\u5316\u95ee\u9898\u7684\u6df1\u5ea6\uff0c\u7b80\u8857\u96c6\u56e2\u5bf9\u5176\u73b0\u6709\u4ea4\u6613\u6a21\u578b\u5728\u8fd9\u4e00\u7279\u5b9a\u95ee\u9898\u4e0a\u7684\u8868\u73b0\u751a\u4e3a\u6ee1\u610f\u3002\u7136\u800c\uff0c\u6ca1\u6709\u4ec0\u4e48\u6bd4\u4e00\u4e2a\u597d\u7684\u8c1c\u9898\u66f4\u51fa\u8272\u7684\u4e86\uff0c\u671f\u671b\u8fd9\u9879\u6311\u6218\u8d5b\u80fd\u591f\u6210\u4e3a\u7b80\u8857\u96c6\u56e2\u6bcf\u65e5\u6216\u8bb8\u8981\u5904\u7406\u7684\u4e00\u79cd\u6570\u636e\u79d1\u5b66\u95ee\u9898\u7684\u6709\u8da3\u5bfc\u8bba\u3002\u7b80\u8857\u96c6\u56e2\u671f\u5f85\u770b\u5230 Kaggle \u793e\u533a\u5c06\u91c7\u7528\u5168\u65b0\u4e14\u5bcc\u6709\u521b\u9020\u6027\u7684\u65b9\u6cd5\u6765\u5e94\u5bf9\u8fd9\u4e00\u4ea4\u6613\u6311\u6218\u3002<\/p>\n\n\n\n
\u8bc4\u4f30\u65b9\u5f0f\uff1a\u672c\u6b21\u6bd4\u8d5b\u4ee5\u6548\u7528\u5f97\u5206\u4f5c\u4e3a\u8861\u91cf\u5c3a\u5ea6\u3002\u6d4b\u8bd5\u96c6\u4e2d\u7684\u6bcf\u4e00\u884c\u4ee3\u8868\u4e00\u4e2a\u4ea4\u6613\u673a\u4f1a\uff0c\u60a8\u9700\u8981\u9884\u6d4b\u8be5\u673a\u4f1a\u7684 action \u4ef7\u503c\uff0c1 \u8868\u793a\u8fdb\u884c\u4ea4\u6613\uff0c0 \u8868\u793a\u653e\u5f03\u4ea4\u6613\u3002\u6bcf\u7b14\u4ea4\u6613 j \u90fd\u6709\u4e00\u4e2a\u76f8\u5173\u8054\u7684 weight \u548c resp\uff0c\u4ee3\u8868\u56de\u62a5\u3002\u5176\u4e2d\uff0c\u9488\u5bf9\u7b2c i \u4e2a\u4ea4\u6613\u65e5\uff0c\u6211\u4eec\u8fdb\u884c\u5b9a\u4e49\uff1a<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\u6700\u540e\u7684\u8bc4\u5206\u8868\u793a\u4e3a\uff1a<\/p>\n\n\n
\n
\"\"<\/figure>\n<\/div>\n\n\n
\u4e8c\u3001\u91d1\u724c\u6a21\u578b\u4ecb\u7ecd<\/h3>\n\n\n\n
\u6a21\u578b\u91c7\u7528AutoEncoder\u4ee5\u53caMLP\u7684\u67b6\u6784\uff0c\u6574\u4f53\u6846\u67b6\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n\n\n
\n
\"\"<\/figure>\n<\/div>\n\n\n
\u5176\u4e2d\uff0cAutoencoder \u90e8\u5206\u8fd0\u7528\u4e86\u53bb\u566a\u81ea\u7f16\u7801\u5668\uff08Denoise Autoencoder\uff09\uff0c\u4e5f\u5c31\u662f\u5728\u8f93\u5165\u6570\u636e\u8fdb\u5165 Encoder \u4e4b\u524d\uff0c\u9996\u5148\u6dfb\u52a0\u9ad8\u65af\u566a\u58f0\uff0c\u5982\u6b64\u53ef\u8d77\u5230\u6570\u636e\u589e\u5f3a\u7684\u6548\u679c\uff0c\u6709\u5229\u4e8e\u7f13\u89e3\u8fc7\u62df\u5408\uff0c\u540c\u65f6\u8fd8\u80fd\u591f\u63d0\u5347\u7279\u5f81\u7684\u8868\u793a\u80fd\u529b\u3002\u81ea\u7f16\u7801\u5668\u7684 Loss \u6db5\u76d6\u4e24\u4e2a\u90e8\u5206\uff0c\u4e00\u90e8\u5206\u662f MSE Loss\uff0c\u7528\u4e8e\u6570\u636e\u91cd\u6784\uff1b\u53e6\u4e00\u90e8\u5206\u662f\u901a\u8fc7\u4e00\u4e2a\u5206\u7c7b\u5668\u6a21\u5757\u5f00\u5c55\u591a\u5206\u7c7b\u4efb\u52a1\uff0c\u4ee5\u6b64\u5b9e\u73b0 action \u7684\u5206\u7c7b\uff0c\u786e\u4fdd Encoder \u5b66\u4e60\u83b7\u53d6\u7684\u7279\u5f81\u5177\u5907\u8ddf\u4ea4\u6613\u52a8\u4f5c\u76f8\u5173\u7684\u7279\u6027\u3002MLP \u90e8\u5206\u7684\u8f93\u5165\u5305\u542b\u4e24\u90e8\u5206\uff0c\u5373\u539f\u59cb\u6570\u636e\u8f93\u5165\u4ee5\u53ca Encoder \u7684\u7f16\u7801\u7ed3\u679c\uff0c\u591a\u5c42 MLP \u5f97\u5230\u7684\u7ed3\u679c\u7528\u4e8e\u6700\u7ec8\u4ea4\u6613\u52a8\u4f5c\u7684\u5206\u7c7b\u3002<\/p>\n\n\n\n
\u8fd9\u90e8\u5206\u7684\u5b9e\u73b0\u4ee3\u7801\u5982\u4e0b\uff0c\u5b8c\u6574\u6e90\u7801\u53ef\u4ee5\u53bb\u8be5notebook\u4e0b\u83b7\u53d6\uff1a<\/p>\n\n\n\n
https:\/\/www.kaggle.com\/code\/gogo827jz\/jane-street-supervised-autoencoder-mlp\/notebook?scriptVersionId=73762661<\/a><\/p>\n\n\n\n
def create_ae_mlp(num_columns, num_labels, hidden_units, dropout_rates, ls = 1e-2, lr = 1e-3):\n    \n    inp = tf.keras.layers.Input(shape = (num_columns, ))\n    x0 = tf.keras.layers.BatchNormalization()(inp)\n    \n    encoder = tf.keras.layers.GaussianNoise(dropout_rates[0])(x0)\n    encoder = tf.keras.layers.Dense(hidden_units[0])(encoder)\n    encoder = tf.keras.layers.BatchNormalization()(encoder)\n    encoder = tf.keras.layers.Activation('swish')(encoder)\n    \n    decoder = tf.keras.layers.Dropout(dropout_rates[1])(encoder)\n    decoder = tf.keras.layers.Dense(num_columns, name = 'decoder')(decoder)\n\n    x_ae = tf.keras.layers.Dense(hidden_units[1])(decoder)\n    x_ae = tf.keras.layers.BatchNormalization()(x_ae)\n    x_ae = tf.keras.layers.Activation('swish')(x_ae)\n    x_ae = tf.keras.layers.Dropout(dropout_rates[2])(x_ae)\n\n    out_ae = tf.keras.layers.Dense(num_labels, activation = 'sigmoid', name = 'ae_action')(x_ae)\n    \n    x = tf.keras.layers.Concatenate()([x0, encoder])\n    x = tf.keras.layers.BatchNormalization()(x)\n    x = tf.keras.layers.Dropout(dropout_rates[3])(x)\n    \n    for i in range(2, len(hidden_units)):\n        x = tf.keras.layers.Dense(hidden_units[i])(x)\n        x = tf.keras.layers.BatchNormalization()(x)\n        x = tf.keras.layers.Activation('swish')(x)\n        x = tf.keras.layers.Dropout(dropout_rates[i + 2])(x)\n        \n    out = tf.keras.layers.Dense(num_labels, activation = 'sigmoid', name = 'action')(x)\n    \n    model = tf.keras.models.Model(inputs = inp, outputs = [decoder, out_ae, out])\n    model.compile(optimizer = tf.keras.optimizers.Adam(learning_rate = lr),\n                  loss = {'decoder': tf.keras.losses.MeanSquaredError(), \n                          'ae_action': tf.keras.losses.BinaryCrossentropy(label_smoothing = ls),\n                          'action': tf.keras.losses.BinaryCrossentropy(label_smoothing = ls), \n                         },\n                  metrics = {'decoder': tf.keras.metrics.MeanAbsoluteError(name = 'MAE'), \n                             'ae_action': tf.keras.metrics.AUC(name = 'AUC'), \n                             'action': tf.keras.metrics.AUC(name = 'AUC'), \n                            }, \n                 )\n    \n    return model<\/code><\/pre>\n\n\n\n
\u6a21\u578b\u8bad\u7ec3\u7684\u65f6\u5019\u91c7\u7528\u4e86PurgedGroupTimeSeriesSplit\u65b9\u5f0f\u5bf9\u8bad\u7ec3\u6570\u636e\u8fdb\u884c\u5212\u5206\uff0c\u5b83\u53ef\u4ee5\u4fdd\u8bc1\u4e0d\u4f1a\u6709\u672a\u6765\u6570\u636e\u7684\u6cc4\u9732\uff0c\u540c\u65f6\u4e5f\u53ef\u4ee5\u91c7\u7528n-folds\u7684\u65b9\u5f0f\u8fdb\u884c\u4ea4\u53c9\u9a8c\u8bc1\uff0c\u5b9e\u9a8c\u65f6\u4f5c\u8005\u91c7\u7528\u4e865-folds\u7684\u65b9\u5f0f\u3002<\/p>\n\n\n\n
if not TEST:
    scores = []
    batch_size = 4096
    gkf = PurgedGroupTimeSeriesSplit(n_splits = n_splits, group_gap = group_gap)
    for fold, (tr, te) in enumerate(gkf.split(train['action'].values, train['action'].values, train['date'].values)):
        ckp_path = f'JSModel_{fold}.hdf5'
        model = create_ae_mlp(**params)
        ckp = ModelCheckpoint(ckp_path, monitor = 'val_action_AUC', verbose = 0, 
                              save_best_only = True, save_weights_only = True, mode = 'max')
        es = EarlyStopping(monitor = 'val_action_AUC', min_delta = 1e-4, patience = 10, mode = 'max', 
                           baseline = None, restore_best_weights = True, verbose = 0)
        history = model.fit(X[tr], [X[tr], y[tr], y[tr]], validation_data = (X[te], [X[te], y[te], y[te]]), 
                            sample_weight = sw[tr], 
                            epochs = 100, batch_size = batch_size, callbacks = [ckp, es], verbose = 0)
        hist = pd.DataFrame(history.history)
        score = hist['val_action_AUC'].max()
        print(f'Fold {fold} ROC AUC:\\t', score)
        scores.append(score)

        K.clear_session()
        del model
        rubbish = gc.collect()
    
    print('Weighted Average CV Score:', weighted_average(scores))<\/code><\/pre>\n\n\n\n
\u9664\u6b64\u4ee5\u5916\uff0c\u4f5c\u8005\u4ea6\u8fd0\u7528\u4e86\u4e00\u4e9b\u522b\u7684 tricks\uff0c\u50cf\u662f Early-Stop\u3001BatchNormalization \u8fd8\u6709\u8d85\u53c2\u6570\u641c\u7d22\u7b49\u7b49\uff0c\u6b64\u5916\uff0c\u81ea\u7f16\u7801\u5668\u7684\u6fc0\u6d3b\u51fd\u6570\u8fd0\u7528\u5230\u4e86 swish\uff0c\u800c\u975e relu \u6291\u6216 leaky-relu\uff0c\u4e0d\u8fc7\u4f9d\u636e\u4f5c\u8005\u6240\u8a00\uff0c\u91c7\u7528\u54ea\u79cd\u6fc0\u6d3b\u51fd\u6570\u5dee\u522b\u4e0d\u5927\u3002\u5982\u679c\u5927\u5bb6\u5bf9\u8fd9\u4e2a\u65b9\u6848\u611f\u5174\u8da3\uff0c\u63a8\u8350\u5ef6\u5c55\u9605\u8bfb\u5982\u4e0b\uff1a<\/p>\n\n\n\n
https:\/\/www.kaggle.com\/competitions\/jane-street-market-prediction\/data<\/a><\/p>\n\n\n\n
https:\/\/www.kaggle.com\/code\/gogo827jz\/jane-street-supervised-autoencoder-mlp\/data?scriptVersionId=73762661&select=JSModel_0.hdf5<\/a><\/p>\n\n\n\n
\n\n\n\n
\u672c\u6587\u5185\u5bb9\u4ec5\u4ec5\u662f\u6280\u672f\u63a2\u8ba8\u548c\u5b66\u4e60\uff0c\u5e76\u4e0d\u6784\u6210\u4efb\u4f55\u6295\u8d44\u5efa\u8bae\u3002<\/mark><\/p>\n","protected":false},"excerpt":{"rendered":"
\u5199\u5728\u524d\u9762\u7684\u8bdd\uff1a\u201c\u4f4e\u4e70\u9ad8\u5356\u201d\u542c\u8d77\u6765\u5f88\u7b80\u5355\u2026\u2026\u4e8b\u5b9e\u4e0a\uff0c…<\/p>\n
Continue reading\u83b7\u5f97\u7b80\u8857\u5e02\u573a\u9884\u6d4b\u5927\u8d5b\u91d1\u724c\u7684\u9884\u6d4b\u7b56\u7565\u6a21\u578b<\/span><\/a><\/div>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[6],"class_list":["post-841","post","type-post","status-publish","format-standard","hentry","category-aiinvest","tag-6","entry"],"_links":{"self":[{"href":"https:\/\/laoyulaoyu.com\/index.php\/wp-json\/wp\/v2\/posts\/841","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/laoyulaoyu.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/laoyulaoyu.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/laoyulaoyu.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/laoyulaoyu.com\/index.php\/wp-json\/wp\/v2\/comments?post=841"}],"version-history":[{"count":0,"href":"https:\/\/laoyulaoyu.com\/index.php\/wp-json\/wp\/v2\/posts\/841\/revisions"}],"wp:attachment":[{"href":"https:\/\/laoyulaoyu.com\/index.php\/wp-json\/wp\/v2\/media?parent=841"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/laoyulaoyu.com\/index.php\/wp-json\/wp\/v2\/categories?post=841"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/laoyulaoyu.com\/index.php\/wp-json\/wp\/v2\/tags?post=841"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}