f1score的Spark mllib阈值

Question

我正在尝试找到最佳阈值，以使我的逻辑回归具有最高的f1分数。但是，当我写下以下几行代码时：

val f1Score = metrics.fMeasureByThreshold
f1Score.foreach { case (t, f) =>
println(s"Threshold: $t, F-score: $f, Beta = 1")

出现了一些奇怪的值，例如：

Threshold: 2.0939996826644833, F-score: 0.285648784961027, Beta = 1
Threshold: 2.093727854652065, F-score: 0.28604171441668574, Beta = 1
Threshold: 2.0904571465313113, F-score: 0.2864344637946838, Beta = 1
Threshold: 2.0884466833553468, F-score: 0.28682703321878583, Beta = 1
Threshold: 2.0882666552407283, F-score: 0.2872194228126431, Beta = 1
Threshold: 2.0835997800203447, F-score: 0.2876116326997939, Beta = 1
Threshold: 2.077892816382506, F-score: 0.28800366300366304, Beta = 1

怎么可能有一个大于1的阈值？负值也是如此，这些负值在控制台输出中进一步显示。

Answer 1

我之前在将Dataframe转换为RDD时犯了一个错误，而不是写：

val  predictionAndLabels =predictions.select("probability", "labelIndex").rdd.map(x => (x(0).asInstanceOf[DenseVector](1), x(1).asInstanceOf[Double]))

我写道：

val  predictionAndLabels =predictions.select("rawPredictions", "labelIndex").rdd.map(x => (x(0).asInstanceOf[DenseVector](1), x(1).asInstanceOf[Double]))

所以阈值是关于rawPredictions的，而不是概率，现在一切都有意义了