Forward Stagewise Algorithm
1. 算法描述
- 输入:
- 训练数据集:\(T={(x_1, y_1), (x_2, y_2), \ldots, (x_n, y_n)}\)
- 损失函数:L(y, f(x))。\(y\):样本标签向量,\(f(x)\):预测结果向量
- 基函数集:\({b(x, \gamma)}\)。\(\gamma\):模型参数向量, 一组\(\gamma\)对应一个子模型
- 输出: 训练M个模型\(b(x, \gamma_m)\),按模型权重\(\beta_m\)相加得到最红加法模型,如下: \[ \boxed{ f(x) = \sum_{i=1}^M \beta_m b(x, \gamma_m) } \]
2. 算法流程:
- \(f_0(x) = 0\)
- for m in 1, 2, \(\ldots\), M:
- \[ \begin{multline} (\beta_m, \gamma) = arg \min_{\beta, \gamma} \sum_{i=1}^m(y_i, f_{m-1}+\beta b(x_i;\gamma) \end{multline} \]
- \[ \begin{multline} f_m(x) = f_{m-1}(x) + \beta_m b(x; \gamma_m) \end{multline} \]
- \(f(x) = f_m(x)\)
3. 参考资料
[1]《统计学习方法》,李航著,2012