不平衡类上的Python XgBoost

Question

我目前正在做一个机器学习项目，旨在预测一个二进制类(负: 0，正: 1)。数据集不平衡。正值的比例为0.1%。

我正在运行一个xgboost模型，使用基尼系数作为性能度量。问题是，在助推迭代期间，它需要大量的运行来提高分数。

例如：

[Fold 1/2]
[0] train-gini:-0.048192    validation-gini:-0.042979
Multiple eval metrics have been passed: 'validation-gini' will be used for early stopping.

Will train until validation-gini hasn't improved in 200 rounds.
[10]    train-gini:-0.048192    validation-gini:-0.042979
[20]    train-gini:-0.048192    validation-gini:-0.042979
[30]    train-gini:-0.048192    validation-gini:-0.042979
[40]    train-gini:-0.048192    validation-gini:-0.042979
[50]    train-gini:-0.048192    validation-gini:-0.042979
[60]    train-gini:-0.048192    validation-gini:-0.042979
[70]    train-gini:-0.048192    validation-gini:-0.042979
[80]    train-gini:-0.048192    validation-gini:-0.042979
[90]    train-gini:0.197521 validation-gini:0.114222
[100]   train-gini:0.247692 validation-gini:0.150601
[110]   train-gini:0.2742   validation-gini:0.169023
[120]   train-gini:0.278983 validation-gini:0.168095
[130]   train-gini:0.316636 validation-gini:0.19118
[140]   train-gini:0.347296 validation-gini:0.191045
[150]   train-gini:0.368581 validation-gini:0.20094
[160]   train-gini:0.374773 validation-gini:0.20906
[170]   train-gini:0.398815 validation-gini:0.215193
[180]   train-gini:0.426088 validation-gini:0.220467
[190]   train-gini:0.439271 validation-gini:0.22249
[200]   train-gini:0.455897 validation-gini:0.226621
[210]   train-gini:0.469989 validation-gini:0.229512
[220]   train-gini:0.485784 validation-gini:0.233432
[230]   train-gini:0.496734 validation-gini:0.23747
[240]   train-gini:0.503718 validation-gini:0.241804
[250]   train-gini:0.51102  validation-gini:0.241841
[260]   train-gini:0.523444 validation-gini:0.244312
[270]   train-gini:0.530968 validation-gini:0.245467
[280]   train-gini:0.538703 validation-gini:0.247433
[290]   train-gini:0.546911 validation-gini:0.244196
[300]   train-gini:0.553623 validation-gini:0.244161
[310]   train-gini:0.561385 validation-gini:0.245099
[320]   train-gini:0.571532 validation-gini:0.244787
[330]   train-gini:0.578088 validation-gini:0.246146
[340]   train-gini:0.585054 validation-gini:0.245624
[350]   train-gini:0.591924 validation-gini:0.245463
[360]   train-gini:0.596331 validation-gini:0.247517
[370]   train-gini:0.600661 validation-gini:0.249465
[380]   train-gini:0.606264 validation-gini:0.249034
[390]   train-gini:0.611768 validation-gini:0.249182
[400]   train-gini:0.617176 validation-gini:0.248239
[410]   train-gini:0.621629 validation-gini:0.249248
[420]   train-gini:0.626766 validation-gini:0.24975
[430]   train-gini:0.631587 validation-gini:0.247824
[440]   train-gini:0.636737 validation-gini:0.246586
[450]   train-gini:0.641735 validation-gini:0.246552
[460]   train-gini:0.649765 validation-gini:0.246332
[470]   train-gini:0.654319 validation-gini:0.243546
[480]   train-gini:0.659301 validation-gini:0.241965
[490]   train-gini:0.665632 validation-gini:0.242562
[500]   train-gini:0.669333 validation-gini:0.241306
[510]   train-gini:0.673625 validation-gini:0.240314
[520]   train-gini:0.678935 validation-gini:0.239846
[530]   train-gini:0.683851 validation-gini:0.240029
[540]   train-gini:0.685694 validation-gini:0.240691
[550]   train-gini:0.689285 validation-gini:0.239974
[560]   train-gini:0.691698 validation-gini:0.239079
[570]   train-gini:0.694017 validation-gini:0.239407
Stopping. Best iteration:
[373]   train-gini:0.60227  validation-gini:0.24996

我们可以看到，在第80轮，训练和验证的分数最终得到了提高。即使我改变了分割的种子，这种情况也会重复(但分数增加的这一轮的n°将改变)。

有没有人遇到过这样的问题？

干杯，Astrus

Answer 1

不是的。但是，如果只有0.1%的正值，您可能需要尝试scale_pos_weight : float参数的xgboost值

也许它能解决这个问题。我会这么说：

scale_pos_weight = 1000

Answer 2

你有没有尝试过按照xgboost documentation将eval_metric更改为logloss或error