data_mining:xgboost

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revisionPrevious revision
Next revision
Previous revision
data_mining:xgboost [2019/05/03 00:52] – [XGBoost] phreazerdata_mining:xgboost [2020/08/02 16:12] (current) phreazer
Line 20: Line 20:
 $$ $$
  
-$F$ is space of functions containing all regression trees +===== Gradient boosting ===== 
-$K$ is number of trees + 
-$f_k(x_i)$ is regression tree that maps a attribute to a score+  * $F$ is space of functions containing all regression trees 
 +  $K$ is number of trees 
 +  $f_k(x_i)$ is regression tree that maps a attribute to a score
  
 Learn functions (trees) instead of weights in $R^d$. Learn functions (trees) instead of weights in $R^d$.
Line 33: Line 35:
  
 Learning objective: Learning objective:
-  * Training loss: Fit of the functions to the points +  * **Training loss**: Fit of the functions to the points 
-  * Regularization: Complexity of function; Number of splitting points, l2 norm of height in each segment+  * **Regularization**: Complexity of function; Number of splitting points, l2 norm of height in each segment
  
 Objective: Objective:
Line 57: Line 59:
   * Logistic loss $l(y_i,\hat{y}_i)=y_i \ln(1+e^{-\hat{y}_i})+(1-y_i)\ln(1+e^{\hat{y}_i})$ (LogitBoost)   * Logistic loss $l(y_i,\hat{y}_i)=y_i \ln(1+e^{-\hat{y}_i})+(1-y_i)\ln(1+e^{\hat{y}_i})$ (LogitBoost)
  
-Stochastic Gradient Descent can not be applied, since trees are used.+Stochastic Gradient descent can not be applied, since trees are used.
  
 Solution is **additive training**: Start with constant prediction, add a new function each time. Solution is **additive training**: Start with constant prediction, add a new function each time.
Line 80: Line 82:
  
  
-Taylor expansion approximation of loss+==== Taylor expansion ====
  
 Use taylor expansion to approximate a function through a power series (polynom). Use taylor expansion to approximate a function through a power series (polynom).
Line 93: Line 95:
 $$\sum^n_{i=1} [l(y_i,\hat{y}_i^{(t-1)}) + g_if_t(x_i) + \frac{1}{2}h_if_t^2(x_i)]$$ with $g_i=\delta_{\hat{y}^{(t-1)}} l(y_i,\hat{y}^{(t-1)})$ and $h_i=\delta^2_{\hat{y}^{(t-1)}} l(y_i,\hat{y}^{(t-1)})$ $$\sum^n_{i=1} [l(y_i,\hat{y}_i^{(t-1)}) + g_if_t(x_i) + \frac{1}{2}h_if_t^2(x_i)]$$ with $g_i=\delta_{\hat{y}^{(t-1)}} l(y_i,\hat{y}^{(t-1)})$ and $h_i=\delta^2_{\hat{y}^{(t-1)}} l(y_i,\hat{y}^{(t-1)})$
  
-With removed constants+With removed constants (and square loss)
 $$\sum^n_{i=1} [g_if_t(x_i) + \frac{1}{2}h_if_t^2(x_i)] + \Omega(f_t)$$  $$\sum^n_{i=1} [g_if_t(x_i) + \frac{1}{2}h_if_t^2(x_i)] + \Omega(f_t)$$ 
 So that learning function only influences $g_i$ and $h_i$ while rest stays the same. So that learning function only influences $g_i$ and $h_i$ while rest stays the same.
  • data_mining/xgboost.1556837550.txt.gz
  • Last modified: 2019/05/03 00:52
  • by phreazer