高级集成学习技巧
Examined ensemble methods
- Averaging (or blending)
- Weighted averaging
- Conditional averaging
- Bagging
- Boosting
- Stacking
- StackNet
Averaging ensemble methods
举个例子,假设我们有一个名为age的变量,就像年龄一样,我们试着预测它。我们有两个模型:
-
低于50,模型效果更好
-
高于50,模型效果更好
那么如果我们试图结合它们将会发生什么呢?
Averaging(or blending)
- (model1 + model2) / 2
\(R^2\)上升到0.95,较之前有所改善。但该模型并没有比单模型做的好的地方更好,尽管如此,它平均表现更好。也许可能会有更好的组合呢?来试试加权平均
Weighted averaging
- (model1 x 0.7 + model 2 x 0.3)
看起来没有之前的好
Conditional averaging
- 各取好的部分
理想情况下,我们希望得到类似的结果
Bagging
Why Bagging
建模中有两个主要误差来源
- 1.由于偏差而存在误差(underfitting)
- 2.由于方差而存在误差(overfitting)
通过略微不同的模型,确保预测不会有读取非常高的方差。这通常使它更具普遍性。
Parameters that control bagging?
- Changing the seed
- Row(Sub) sampling or Bootstrapping
- Shuffling
- Column(Sub) sampling
- Model-specific parameters
- Number of models (or bags)
- (Optionally) parallelism
Examples of bagging
Boosting
Boosting是对每个模型构建的模型进行加权平均的一种形式,顺序地考虑以前的模型性能。
Weight based boosting
假设我们有一个表格数据集,有四个特征。 我们称它们为x0,x1,x2和x3,我们希望使用这些功能来预测目标变量y。
我们将预测值称为pred,这些预测有一定的误差。我们可以计算这些绝对误差,|y - pred|
。我们可以基于此生成一个新列或向量,在这里我们创建一个权重列,使用1加上绝对误差。当然有不同的方法来计算这个权重,现在我们只是以此为例。
所有接下来要做的是用这些特征去拟合新的模型,但每次也要增加这个权重列。这就是按顺序添加模型的方法。
Weight based boosting parameters
- Learning rate (or shrinkage or eta)
- 每个模型只相信一点点:
predictionN = pred0*eta + pred1*eta + ... + predN*eta
- Number of estimators
- estimators扩大一倍,eta减小一倍
- Input model - can be anything that accepts weights
- Sub boosting type:
- AdaBoost-Good implementation in sklearn(python)
- LogitBoost-Good implementation in Weka(Java)
Residual based boosting [&]
我们使用同样的数据集做相同的事。预测出pred后
接下来会计算误差
将error作为新的y得到新的预测new_pred
以Rownum=1为例:
最终预测=0.75 + 0.20 = 0.95更接近于1
这种方法很有效,可以很好的减小误差。
Residual based boosting parameters
- Learning rate (or shrinkage or eta)
predictionN = pred0 + pred1*eta + ... + predN*eta
- 前面的例子,如果eta为0.1,则Prediction=0.75 + 0.2*(0.1) = 0.77
- Number of estimators
- Row (sub)sampling
- Column (sub)sampling
- Input model - better be trees.
- Sub boosting type:
- Full gradient based
- Dart
Residual based favourite implementations
- Xgboost
- Lightgbm
- H2O's GBM
- Catboost
- Sklearn's GBM
Stacking
Methodology
- Wolpert in 1992 introduced stacking. It involves:
-
- Splitting the train set into two disjoint sets.
-
- Train several base learners on the first part.
-
- Make predictions with the base learners on the second (validation) part.
具体步骤
假设有A,B,C三个数据集,其中A,B的目标变量y已知。
然后
-
算法0拟合A,预测B和C,然后保存pred0到B1,C1
-
算法1拟合A,预测B和C,然后保存pred1到B1,C1
-
算法2拟合A,预测B和C,然后保存pred2到B1,C1
-
算法3拟合B1,预测C1,得到最终结果preds3
Stacking example
from sklearn.ensemble import RandomForestRegressor
from sklearn.linear_model import LinearRegression
import numpy as np
from sklearn.model_selection import train_test_split
train = '' # your training set
y = '' # your target variable
# split train data in 2 part, training and valdiation.
training, valid, ytraining, yvalid = train_test_split(train, y, test_size=0.5)
# specify models
model1 = RandomForestRegressor()
model2 = LinearRegression()
#fit models
model1.fit(training, ytraining)
model2.fit(trainging, ytraining)
# make predictions for validation
preds1 = model1.predict(valid)
preds2 = model2.predict(valid)
# make predictions for test data
test_preds1 = model1.predict(test)
test_preds2 = model2.predict(test)
# From a new dataset for valid and test via stacking the predictions
stacked_predictions = np.colum_stack((preds1, preds2))
stacked_test_predictions = np.column_stack((test_preds1, test_preds2))
# specify meta model
meta_model = LinearRegression()
meta_model.fit(stacked_predictions, yvalid)
# make predictions on the stacked predictions of the test data
final_predictions = meta_model.predict(stacked_test_predictions)
Stacking(past) example
可以看到,它与我们使用Conditional averaging
的结果非常近似。只是在50附件做的不够好,这是有道理的,因为模型没有见到目标变量,无法准确识别出50这个缺口。所以它只是尝试根据模型的输入来确定。
Things to be mindful of
- With time sensitive data - respect time
- 如果你的数据带有时间元素,你需要指定你的stacking,以便尊重时间。
- Diversity as important as performance
- 单一模型表现很重要,但模型的多样性也非常重要。当模型是坏的或弱的情况,你不需太担心,stacking实际上可以从每个预测中提取到精华,得到好的结果。因此,你真正需要关注的是,我正在制作的模型能给我带来哪些信息,即使它通常很弱。
- Diversity may come from:
- Different algorithms
- Different input features
- Performance plateauing after N models
- Meta model is normally modest
StackNet
https://github.com/kaz-Anova/StackNet
Ensembling Tips and Tricks
\(1^{st}\) level tips
- Diversity based on algorithms:
- 2-3 gradient boosted trees (lightgbm, xgboost, H2O, catboost)
- 2-3 Neural nets (keras, pytorch)
- 1-2 ExtraTrees/RandomForest (sklearn)
- 1-2 linear models as in logistic/ridge regression, linear svm (sklearn)
- 1-2 knn models (sklearn)
- 1 Factorization machine (libfm)
- 1 svm with nonlinear kernel(like RBF) if size/memory allows (sklearn)
- Diversity based on input data:
- Categorical features: One hot, label encoding, target encoding, likelihood encoding, frequency or counts
- Numerical features: outliers, binning, derivatives, percentiles, scaling
- Interactions: col1*/+-col2, groupby, unsupervised
\(2^{st}\) level tips
-
Simpler (or shallower) Algorithms:
-
gradient boosted trees with small depth(like 2 or 3)
-
Linear models with high regularization
-
Extra Trees (just don't make them too big)
-
Shallow networks (as in 1 hidden layer, with not that many hidden neurons)
-
knn with BrayCurtis Distance
-
Brute forcing a search for best linear weights based on cv
-
Feature engineering:
-
pairwise differences between meta features
-
row-wise statistics like averages or stds
-
Standard feature selection techniques
-
For every 7.5 models in previous level we add 1 in meta (经验)
-
Be mindful to target leakage