簡介
NDCG是衡量Ranking quality是重要指標。本文將用實際Python例子演示怎樣計算NDCG。
用scikit計算NDCG例子
注意ndcg_score接收的參數都是list of list。後面解釋為什麼是list of list。
如果評估一個排序請求，用以下例子
from sklearn.metrics import ndcg_score
y_truth = [0, 1, 2, 3]
y_predict = [0, 1, 2, 3]
ndcg = ndcg_score([y_truth], [y_predict])
print(f'ndcg = {ndcg}')
from sklearn.metrics import ndcg_score
y_truth = [0, 1, 2, 3]
y_predict = [0, 1, 2, 3]
ndcg = ndcg_score([y_truth], [y_predict])
print(f'ndcg = {ndcg}')
from sklearn.metrics import ndcg_score
y_truth = [0, 1, 2, 3]
y_predict = [0, 1, 2, 3]
ndcg = ndcg_score([y_truth], [y_predict])
print(f'ndcg = {ndcg}')
如果有多個排序請求，比如日誌里記錄的一天收到的所有排序請求，用以下例子（假設有3個排序請求）
from sklearn.metrics import ndcg_score
y_truth1 = [0, 1, 2, 3]
y_predict2 = [0, 1, 2, 3]
y_truth2 = [0, 1, 2, 3]
y_predict2 = [0, 1, 2, 3]
y_truth3 = [0, 1, 2, 3]
y_predict3 = [0, 1, 2, 3]
ndcg = ndcg_score(
[y_truth1, y_truth2, y_truth3],
[y_predict1, y_predict2, y_predict3]
)
print(f'ndcg = {ndcg}')
from sklearn.metrics import ndcg_score
y_truth1 = [0, 1, 2, 3]
y_predict2 = [0, 1, 2, 3]
y_truth2 = [0, 1, 2, 3]
y_predict2 = [0, 1, 2, 3]
y_truth3 = [0, 1, 2, 3]
y_predict3 = [0, 1, 2, 3]
ndcg = ndcg_score(
[y_truth1, y_truth2, y_truth3],
[y_predict1, y_predict2, y_predict3]
)
print(f'ndcg = {ndcg}')
from sklearn.metrics import ndcg_score
y_truth1 = [0, 1, 2, 3]
y_predict2 = [0, 1, 2, 3]
y_truth2 = [0, 1, 2, 3]
y_predict2 = [0, 1, 2, 3]
y_truth3 = [0, 1, 2, 3]
y_predict3 = [0, 1, 2, 3]
ndcg = ndcg_score(
[y_truth1, y_truth2, y_truth3],
[y_predict1, y_predict2, y_predict3]
)
print(f'ndcg = {ndcg}')
3個排序會產生3個NDCG，最後結果為所有NDCG的平均值，作為這麼多請求的總NDCG返回。
還要注意每個排序請求list裡面元素的數量要一樣，不然ndcg_score函數會報錯。
參考
- Scikit-learn官方文檔 （講得沒本文詳細）
