使用 xgboost 对 Tweedie 回归建模

Question

我试图用来xgboost制作花呢模型，但是我收到一条模糊的错误消息。

这是一个可重现的示例：

准备数据：

library(xgboost)
library(dplyr)

set.seed(123)
xx <- rpois(5000, 0.02)
xx[xx>0] <- rgamma(sum(xx>0), 50)

yy <- matrix(rnorm(15000), 5000,3, dimnames = list(1:5000, c("a", "b", "c")))

train_test <- sample(c(0,1), 5000, replace = T)

准备xgboost，这里最重要的是：objective = 'reg:tweedie'，eval_metric = "tweedie-nloglik"和tweedie_variance_power = 1.2：

dtrain <- xgb.DMatrix(
  data = yy %>% subset(train_test == 0),
  label = xx %>% subset(train_test == 0)
)

dtest <- xgb.DMatrix(
  data = yy %>% subset(train_test == 1),
  label = xx %>% subset(train_test == 1)
)

watchlist <- list(eval = dtest, train = dtrain)

param <- list(max.depth = 2,
              eta = 0.3,
              nthread = 1,
              silent = 1,
              objective = 'reg:tweedie',
              eval_metric = "tweedie-nloglik",
              tweedie_variance_power = 1.2)

最后调用 xgboost：

resBoost <- xgb.train(params = param, data=dtrain, nrounds = 20, watchlist=watchlist)

这给出了这个晦涩的错误信息：

Error in xgb.iter.update(bst$handle, dtrain, iteration - 1, obj) :
  [17:59:18] amalgamation/../src/metric/elementwise_metric.cc:168: Check failed: param != nullptr tweedie-nloglik must be in formattweedie-nloglik@rho

Stack trace returned 10 entries:
[bt] (0) /usr/local/lib/R/site-library/xgboost/libs/xgboost.so(dmlc::StackTrace[abi:cxx11]()+0x1bc) [0x7f1f0ce742ac]
[bt] (1) /usr/local/lib/R/site-library/xgboost/libs/xgboost.so(dmlc::LogMessageFatal::~LogMessageFatal()+0x28) [0x7f1f0ce74e88]
[bt] (2) /usr/local/lib/R/site-library/xgboost/libs/xgboost.so(xgboost::metric::EvalTweedieNLogLik::EvalTweedieNLogLik(char const*)+0x1eb) [0x7f1f0cea00db]
[bt] (3) /usr/local/lib/R/site-library/xgboost/libs/xgboost.so(+0x68ef1) [0x7f1f0ce78ef1]
[bt] (4) /usr/local/lib/R/site-library/xgboost/libs/xgboost.so(xgboost::Metric::Create(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&)+0x263) [0x7f1f0ce7ede3]
[bt] (5) /usr/local/lib/R/site-library/xgboost/libs/xgboost.so(xgboost::LearnerImpl::Configure(std::vector<std::pair

问题似乎与参数有关，eval_metric = "tweedie-nloglik"因为如果我更改eval_metric为logloss它通过：

param$eval_metric <- "logloss"
resBoost <- xgb.train(params = param, data=dtrain, nrounds = 20, watchlist=watchlist)
[1]     eval-logloss:0.634391   train-logloss:0.849734
[2]     eval-logloss:0.634391   train-logloss:0.849734
...

知道如何使用eval_metric = "tweedie-nloglik"参数，因为它在我的上下文中似乎最合适吗？谢谢

Answer 1

TL;DR：感谢Frans Rodenburg评论：use eval_metric="tweedie-nloglik@1.2

我正在查看 tweedie eval 的实现（我什至不知道 tweedie 是什么）以及以下链接中的 logloss eval

花呢：

struct EvalTweedieNLogLik: public EvalEWiseBase<EvalTweedieNLogLik> {
  explicit EvalTweedieNLogLik(const char* param) {
    CHECK(param != nullptr)
        << "tweedie-nloglik must be in format tweedie-nloglik@rho";
    rho_ = atof(param);
    CHECK(rho_ < 2 && rho_ >= 1)
        << "tweedie variance power must be in interval [1, 2)";
    std::ostringstream os;
    os << "tweedie-nloglik@" << rho_;
    name_ = os.str();
  }
  const char *Name() const override {
    return name_.c_str();
  }
  inline bst_float EvalRow(bst_float y, bst_float p) const {
    bst_float a = y * std::exp((1 - rho_) * std::log(p)) / (1 - rho_);
    bst_float b = std::exp((2 - rho_) * std::log(p)) / (2 - rho_);
    return -a + b;
  }
 protected:
  std::string name_;
  bst_float rho_;
};

对数损失：

struct EvalLogLoss : public EvalEWiseBase<EvalLogLoss> {
  const char *Name() const override {
    return "logloss";
  }
  inline bst_float EvalRow(bst_float y, bst_float py) const {
    const bst_float eps = 1e-16f;
    const bst_float pneg = 1.0f - py;
    if (py < eps) {
      return -y * std::log(eps) - (1.0f - y)  * std::log(1.0f - eps);
    } else if (pneg < eps) {
      return -y * std::log(1.0f - eps) - (1.0f - y)  * std::log(eps);
    } else {
      return -y * std::log(py) - (1.0f - y) * std::log(pneg);
    }
  }
};

看起来EvalTweedieNLogLik应该得到一个名为param. 也看起来像你得到了那些确切的行：

CHECK(param != nullptr)
    << "tweedie-nloglik must be in format tweedie-nloglik@rho";

当我将它与EvalLogLoss进行比较时，相关性差异在于它不需要变量，这就是它起作用的原因。

感谢@Frans Rodenburg 评论，我一直在搜索并阅读了如何在此处使用它的示例。

用 eval_metric="tweedie-nloglik@1.2

从 xgboost 文档中阅读这些行时，我第一次也弄错了：

tweedie-nloglik：Tweedie 回归的负对数似然（在 tweedie_variance_power 参数的指定值下）

它可能只与 python 相关。