{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# Datawhale智慧海洋建设-Task4模型建立" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "此部分为智慧海洋建设竞赛的模型建立模块。在该模块中主要介绍了如何进行模型建立并对模型调优。" ] }, { "cell_type": "markdown", "metadata": { "hide_input": true }, "source": [ "## 学习目标\n", "1. 学习如何选择合适的模型以及如何通过模型来进行特征选择\n", "2. 掌握随机森林、lightGBM、Xgboost模型的使用。\n", "3. 掌握贝叶斯优化方法的具体使用" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 内容介绍\n", "1. 模型训练与预测\n", " - 随机森林\n", " - lightGBM模型\n", " - Xgboost模型\n", "2. 交叉验证\n", "3. 模型调参\n", "4. 智慧海洋数据集模型代码示例" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "# 模型训练与预测" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "模型训练与预测的主要步骤为： \n", "(1):导入需要的工具库 \n", "(2):对数据预处理，包括导入数据集、处理数据等操作，具体为缺失值处理、连续特征归一化、类别特征转换等 \n", "(3):训练模型。选择合适的机器学习模型，利用训练集对模型进行训练，达到最佳拟合效果。 \n", "(4):预测结果。将待预测的数据输入到训练好的模型中，得到预测的结果。\n", "\n", "下面进行几种常用的分类算法进行介绍" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 随机森林分类\n", "[随机森林参数介绍](https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html#sklearn.ensemble.RandomForestClassifier) \n", "随机森林是通过集成学习的思想将多棵树集成的一种算法，基本单元是决策树，而它的本质属于机器学习的一个分支——集成学习。\n", "随机森林模型的主要优点是：在当前算法中，具有较好的准确率；能够有效地运行在大数据集上；能够处理具有高维特征的输入样本，而且不需要降维；能够评估各个特征在分类问题上的重要性；在生成过程中，能够获取到内部生成误差的一种无偏估计；对于缺省值问题也能够获得很好的结果。 \n", "\n", "使用sklearn调用随机森林分类树进行预测算法：" ] }, { "cell_type": "code", "execution_count": 19, "metadata": {}, "outputs": [], "source": [ "from sklearn import datasets\n", "from sklearn.ensemble import RandomForestClassifier\n", "from sklearn.model_selection import train_test_split\n", "from sklearn.metrics import f1_score\n", "#数据集导入\n", "iris=datasets.load_iris()\n", "feature=iris.feature_names\n", "X = iris.data\n", "y = iris.target\n", "#随机森林\n", "clf=RandomForestClassifier(n_estimators=200)\n", "train_X,test_X,train_y,test_y = train_test_split(X,y,test_size=0.1,random_state=5)\n", "clf.fit(train_X,train_y)\n", "test_pred=clf.predict(test_X)" ] }, { "cell_type": "code", "execution_count": 38, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "['sepal length (cm)', 'sepal width (cm)', 'petal length (cm)', 'petal width (cm)']\n", "[0.09838896 0.01544017 0.34365936 0.5425115 ]\n" ] } ], "source": [ "#特征的重要性查看\n", "print(str(feature)+'\\n'+str(clf.feature_importances_))" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "采用F1 score进行模型的评价，[此为一篇csdn中对该评价方法的简单说明](https://blog.csdn.net/qq_14997473/article/details/82684300)" ] }, { "cell_type": "code", "execution_count": 13, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "随机森林-macro： 0.818181818181818\n", "随机森林-weighted： 0.8\n" ] } ], "source": [ "#F1-score 用于模型评价\n", "#如果是二分类问题则选择参数‘binary’\n", "#如果考虑类别的不平衡性，需要计算类别的加权平均，则使用‘weighted’\n", "#如果不考虑类别的不平衡性，计算宏平均，则使用‘macro’\n", "score=f1_score(test_y,test_pred,average='macro')\n", "print(\"随机森林-macro：\",score)\n", "score=f1_score(test_y,test_pred,average='weighted')\n", "print(\"随机森林-weighted：\",score)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## lightGBM模型\n", "[lightGBM的学习可参见这篇文章](https://mp.weixin.qq.com/s/64xfT9WIgF3yEExpSxyshQ) \n", "[lightGBM中文文档](https://lightgbm.apachecn.org/#/)这个对超参数的讲解较为详细，建议仔细阅读" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "1. lightGBM过拟合处理方案：\n", "- 使用较小的 max_bin\n", "- 使用较小的 num_leaves\n", "- 使用 min_data_in_leaf 和 min_sum_hessian_in_leaf\n", "- 通过设置 bagging_fraction 和 bagging_freq 来使用 bagging\n", "- 通过设置 feature_fraction 来使用特征子抽样\n", "- 使用更大的训练数据\n", "- 使用 lambda_l1, lambda_l2 和 min_gain_to_split 来使用正则\n", "- 尝试 max_depth 来避免生成过深的树 \n", "2. lightGBM针对更快的训练速度的解决方案\n", "- 通过设置 bagging_fraction 和 bagging_freq 参数来使用 bagging 方法\n", "- 通过设置 feature_fraction 参数来使用特征的子抽样\n", "- 使用较小的 max_bin\n", "- 使用 save_binary 在未来的学习过程对数据加载进行加速\n", "- 使用并行学习, 可参考并行学习指南\n", "3. lightGBM针对更好的准确率\n", "- 使用较大的 max_bin （学习速度可能变慢）\n", "- 使用较小的 learning_rate 和较大的 num_iterations\n", "- 使用较大的 num_leaves （可能导致过拟合）\n", "- 使用更大的训练数据\n", "- 尝试 dart" ] }, { "cell_type": "code", "execution_count": 114, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "[1]\ttrain's multi_logloss: 0.975702\tvalidate's multi_logloss: 1.009\n", "[2]\ttrain's multi_logloss: 0.877457\tvalidate's multi_logloss: 0.914377\n", "[3]\ttrain's multi_logloss: 0.794798\tvalidate's multi_logloss: 0.824134\n", "[4]\ttrain's multi_logloss: 0.723326\tvalidate's multi_logloss: 0.750893\n", "[5]\ttrain's multi_logloss: 0.661667\tvalidate's multi_logloss: 0.682191\n", "[6]\ttrain's multi_logloss: 0.607721\tvalidate's multi_logloss: 0.628136\n", "[7]\ttrain's multi_logloss: 0.560519\tvalidate's multi_logloss: 0.574289\n", "[8]\ttrain's multi_logloss: 0.518687\tvalidate's multi_logloss: 0.529814\n", "[9]\ttrain's multi_logloss: 0.481561\tvalidate's multi_logloss: 0.485778\n", "[10]\ttrain's multi_logloss: 0.448635\tvalidate's multi_logloss: 0.449967\n", "[11]\ttrain's multi_logloss: 0.418864\tvalidate's multi_logloss: 0.414047\n", "[12]\ttrain's multi_logloss: 0.392319\tvalidate's multi_logloss: 0.386407\n", "[13]\ttrain's multi_logloss: 0.368389\tvalidate's multi_logloss: 0.357079\n", "[14]\ttrain's multi_logloss: 0.346782\tvalidate's multi_logloss: 0.33318\n", "[15]\ttrain's multi_logloss: 0.327196\tvalidate's multi_logloss: 0.308858\n", "[16]\ttrain's multi_logloss: 0.309539\tvalidate's multi_logloss: 0.288072\n", "[17]\ttrain's multi_logloss: 0.293482\tvalidate's multi_logloss: 0.268706\n", "[18]\ttrain's multi_logloss: 0.278991\tvalidate's multi_logloss: 0.25158\n", "[19]\ttrain's multi_logloss: 0.265753\tvalidate's multi_logloss: 0.23781\n", "[20]\ttrain's multi_logloss: 0.253744\tvalidate's multi_logloss: 0.226251\n", "[21]\ttrain's multi_logloss: 0.242663\tvalidate's multi_logloss: 0.211902\n", "[22]\ttrain's multi_logloss: 0.232649\tvalidate's multi_logloss: 0.202371\n", "[23]\ttrain's multi_logloss: 0.223439\tvalidate's multi_logloss: 0.192921\n", "[24]\ttrain's multi_logloss: 0.215001\tvalidate's multi_logloss: 0.182008\n", "[25]\ttrain's multi_logloss: 0.2072\tvalidate's multi_logloss: 0.175881\n", "[26]\ttrain's multi_logloss: 0.200111\tvalidate's multi_logloss: 0.168868\n", "[27]\ttrain's multi_logloss: 0.193543\tvalidate's multi_logloss: 0.160918\n", "[28]\ttrain's multi_logloss: 0.187559\tvalidate's multi_logloss: 0.155117\n", "[29]\ttrain's multi_logloss: 0.182121\tvalidate's multi_logloss: 0.148551\n", "[30]\ttrain's multi_logloss: 0.177063\tvalidate's multi_logloss: 0.141508\n", "[31]\ttrain's multi_logloss: 0.172155\tvalidate's multi_logloss: 0.136823\n", "[32]\ttrain's multi_logloss: 0.167851\tvalidate's multi_logloss: 0.13318\n", "[33]\ttrain's multi_logloss: 0.163832\tvalidate's multi_logloss: 0.127932\n", "[34]\ttrain's multi_logloss: 0.160045\tvalidate's multi_logloss: 0.124999\n", "[35]\ttrain's multi_logloss: 0.156511\tvalidate's multi_logloss: 0.11994\n", "[36]\ttrain's multi_logloss: 0.153185\tvalidate's multi_logloss: 0.117388\n", "[37]\ttrain's multi_logloss: 0.150086\tvalidate's multi_logloss: 0.113542\n", "[38]\ttrain's multi_logloss: 0.147138\tvalidate's multi_logloss: 0.11118\n", "[39]\ttrain's multi_logloss: 0.144376\tvalidate's multi_logloss: 0.107657\n", "[40]\ttrain's multi_logloss: 0.141792\tvalidate's multi_logloss: 0.105666\n", "[41]\ttrain's multi_logloss: 0.139327\tvalidate's multi_logloss: 0.102515\n", "[42]\ttrain's multi_logloss: 0.137023\tvalidate's multi_logloss: 0.101176\n", "[43]\ttrain's multi_logloss: 0.134844\tvalidate's multi_logloss: 0.0975092\n", "[44]\ttrain's multi_logloss: 0.132768\tvalidate's multi_logloss: 0.0948682\n", "[45]\ttrain's multi_logloss: 0.130798\tvalidate's multi_logloss: 0.0939896\n", "[46]\ttrain's multi_logloss: 0.128917\tvalidate's multi_logloss: 0.0915695\n", "[47]\ttrain's multi_logloss: 0.127132\tvalidate's multi_logloss: 0.0906398\n", "[48]\ttrain's multi_logloss: 0.12546\tvalidate's multi_logloss: 0.0892012\n", "[49]\ttrain's multi_logloss: 0.123835\tvalidate's multi_logloss: 0.0884964\n", "[50]\ttrain's multi_logloss: 0.122284\tvalidate's multi_logloss: 0.087185\n", "[51]\ttrain's multi_logloss: 0.120772\tvalidate's multi_logloss: 0.0849336\n", "[52]\ttrain's multi_logloss: 0.119346\tvalidate's multi_logloss: 0.0835437\n", "[53]\ttrain's multi_logloss: 0.11795\tvalidate's multi_logloss: 0.0829754\n", "[54]\ttrain's multi_logloss: 0.116534\tvalidate's multi_logloss: 0.0819892\n", "[55]\ttrain's multi_logloss: 0.115189\tvalidate's multi_logloss: 0.0808175\n", "[56]\ttrain's multi_logloss: 0.113915\tvalidate's multi_logloss: 0.0791856\n", "[57]\ttrain's multi_logloss: 0.112663\tvalidate's multi_logloss: 0.0778838\n", "[58]\ttrain's multi_logloss: 0.111477\tvalidate's multi_logloss: 0.0767819\n", "[59]\ttrain's multi_logloss: 0.110319\tvalidate's multi_logloss: 0.0761175\n", "[60]\ttrain's multi_logloss: 0.109189\tvalidate's multi_logloss: 0.075811\n", "[61]\ttrain's multi_logloss: 0.108108\tvalidate's multi_logloss: 0.0743217\n", "[62]\ttrain's multi_logloss: 0.107049\tvalidate's multi_logloss: 0.0730824\n", "[63]\ttrain's multi_logloss: 0.106037\tvalidate's multi_logloss: 0.0725497\n", "[64]\ttrain's multi_logloss: 0.105039\tvalidate's multi_logloss: 0.0709544\n", "[65]\ttrain's multi_logloss: 0.104078\tvalidate's multi_logloss: 0.0703405\n", "[66]\ttrain's multi_logloss: 0.103134\tvalidate's multi_logloss: 0.0701205\n", "[67]\ttrain's multi_logloss: 0.102229\tvalidate's multi_logloss: 0.0692772\n", "[68]\ttrain's multi_logloss: 0.101354\tvalidate's multi_logloss: 0.068559\n", "[69]\ttrain's multi_logloss: 0.100491\tvalidate's multi_logloss: 0.0673473\n", "[70]\ttrain's multi_logloss: 0.0995573\tvalidate's multi_logloss: 0.0674286\n", "[71]\ttrain's multi_logloss: 0.0986634\tvalidate's multi_logloss: 0.0674853\n", "[72]\ttrain's multi_logloss: 0.0978101\tvalidate's multi_logloss: 0.0672873\n", "[73]\ttrain's multi_logloss: 0.0969953\tvalidate's multi_logloss: 0.0673621\n", "[74]\ttrain's multi_logloss: 0.0962072\tvalidate's multi_logloss: 0.066834\n", "[75]\ttrain's multi_logloss: 0.0954358\tvalidate's multi_logloss: 0.06728\n", "[76]\ttrain's multi_logloss: 0.0946999\tvalidate's multi_logloss: 0.0666785\n", "[77]\ttrain's multi_logloss: 0.093984\tvalidate's multi_logloss: 0.0652261\n", "[78]\ttrain's multi_logloss: 0.093268\tvalidate's multi_logloss: 0.0653247\n", "[79]\ttrain's multi_logloss: 0.0925889\tvalidate's multi_logloss: 0.0654675\n", "[80]\ttrain's multi_logloss: 0.0919186\tvalidate's multi_logloss: 0.0649799\n", "[81]\ttrain's multi_logloss: 0.0912796\tvalidate's multi_logloss: 0.0638035\n", "[82]\ttrain's multi_logloss: 0.0906195\tvalidate's multi_logloss: 0.0638154\n", "[83]\ttrain's multi_logloss: 0.0899888\tvalidate's multi_logloss: 0.0642833\n", "[84]\ttrain's multi_logloss: 0.0893663\tvalidate's multi_logloss: 0.0636025\n", "[85]\ttrain's multi_logloss: 0.0887785\tvalidate's multi_logloss: 0.0626043\n", "[86]\ttrain's multi_logloss: 0.0881855\tvalidate's multi_logloss: 0.0623685\n", "[87]\ttrain's multi_logloss: 0.0875885\tvalidate's multi_logloss: 0.0627226\n", "[88]\ttrain's multi_logloss: 0.0870171\tvalidate's multi_logloss: 0.0624081\n", "[89]\ttrain's multi_logloss: 0.0864743\tvalidate's multi_logloss: 0.0625911\n", "[90]\ttrain's multi_logloss: 0.0859347\tvalidate's multi_logloss: 0.0620309\n", "[91]\ttrain's multi_logloss: 0.0854179\tvalidate's multi_logloss: 0.0622157\n", "[92]\ttrain's multi_logloss: 0.0849131\tvalidate's multi_logloss: 0.0617822\n", "[93]\ttrain's multi_logloss: 0.0844192\tvalidate's multi_logloss: 0.0619947\n", "[94]\ttrain's multi_logloss: 0.0839399\tvalidate's multi_logloss: 0.0614539\n", "[95]\ttrain's multi_logloss: 0.0834681\tvalidate's multi_logloss: 0.0616655\n", "[96]\ttrain's multi_logloss: 0.0830149\tvalidate's multi_logloss: 0.06134\n", "[97]\ttrain's multi_logloss: 0.0825657\tvalidate's multi_logloss: 0.0613612\n", "[98]\ttrain's multi_logloss: 0.0821295\tvalidate's multi_logloss: 0.0611025\n", "[99]\ttrain's multi_logloss: 0.0816869\tvalidate's multi_logloss: 0.0613398\n", "[100]\ttrain's multi_logloss: 0.0812595\tvalidate's multi_logloss: 0.0610704\n", "0.9777777777777777\n", "训练集 0.9717813051146384\n", "验证集 0.9734471313418682\n" ] } ], "source": [ "import lightgbm as lgb\n", "from sklearn import datasets\n", "from sklearn.model_selection import train_test_split\n", "import numpy as np\n", "from sklearn.metrics import roc_auc_score, accuracy_score\n", "import matplotlib.pyplot as plt\n", "\n", "# 加载数据\n", "iris = datasets.load_iris()\n", "# 划分训练集和测试集\n", "X_train, X_test, y_train, y_test = train_test_split(iris.data, iris.target, test_size=0.3)\n", "# 转换为Dataset数据格式\n", "train_data = lgb.Dataset(X_train, label=y_train)\n", "validation_data = lgb.Dataset(X_test, label=y_test)\n", "# 参数\n", "results = {}\n", "params = {\n", " 'learning_rate': 0.1,\n", " 'lambda_l1': 0.1,\n", " 'lambda_l2': 0.9,\n", " 'max_depth': 1,\n", " 'objective': 'multiclass', # 目标函数\n", " 'num_class': 3,\n", " 'verbose': -1 \n", "}\n", "# 模型训练\n", "gbm = lgb.train(params, train_data, valid_sets=(validation_data,train_data),valid_names=('validate','train'),evals_result= results)\n", "# 模型预测\n", "y_pred_test = gbm.predict(X_test)\n", "y_pred_data = gbm.predict(X_train)\n", "y_pred_data = [list(x).index(max(x)) for x in y_pred_data]\n", "y_pred_test = [list(x).index(max(x)) for x in y_pred_test]\n", "# 模型评估\n", "print(accuracy_score(y_test, y_pred_test))\n", "print('训练集',f1_score(y_train, y_pred_data,average='macro'))\n", "print('验证集',f1_score(y_test, y_pred_test,average='macro'))" ] }, { "cell_type": "code", "execution_count": 110, "metadata": {}, "outputs": [ { "data": { "image/png": "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\n", "text/plain": [ "

\"img\"