Ikibazo gihuza ibibazo · Hagati Urwego

Imirongo ya ROC hamwe n'ikibazo cya AUC

Reba icyo ROC-AUC ifata, uko ifitanye isano nu rutonde, n'impamvu kalibrasi cyangwa ubusumbane bikeneye isesengura ryihariye.

Inzira zijyanye no kuyoboraImirongo ya Roc na Auc
Ikibazo 1 Bya 8

Ni ubuhe buryo bubiri umugozi wa ROC uteganya?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. Ni ubuhe buryo bubiri umugozi wa ROC uteganya?
    • Igipimo Cyiza Cyukuri Igipimo Cyibinyoma
    • Icyitonderwa na Vuga
    • Ukuri na Threshold
    • Kumva neza na Precision
  2. Nuwuhe gaciro AUC ihuye na classifier idakora neza kuruta gukekeranya?
    • 0.0
    • 0.5
    • 0.75
    • 1.0
  3. AUC irashobora gusobanurwa nkibishoboka ko icyitegererezo:
    • Itondekanya neza icyitegererezo cyatoranijwe
    • Kugera kuri 90% byukuri
    • Amanota meza aruta ayandi adasanzwe
    • Ifite ibyiza bya zeru
  4. Kuki AUC isobanurwa nk '' ubwigenge-bwigenga '?
    • Ikora gusa kuri 0.5
    • Irengagiza icyiciro cyiza rwose
    • Irasaba kalibibishoboka
    • Ihuza imikorere murwego rwose rushoboka
  5. Kuri dataset iringaniye cyane, ni uwuhe murongo usabwa kuruhande cyangwa aho kuba ROC kuko ROC ishobora kugaragara neza?
    • Icyitonderwa-Ibuka umurongo
    • Kwiga umurongo
    • Calibration curve
    • Gutakaza umurongo
  6. Ingingo mu gice cyo hejuru-ibumoso bwa ROC umwanya uhagarariye:
    • Ibyiza byinshi byo kubeshya nibyiza byo hasi
    • Gutandukana neza cyangwa hafi-gutungana
    • Imikorere isanzwe
    • Icyitegererezo kitari cyo
  7. Ni ukubera iki icyitegererezo gishobora kugira ROC-AUC ndende ariko ikigereranyo kidashoboka?
    • AUC directly measures the distance between every predicted probability and its true label
    • Calibration is guaranteed whenever the ROC curve is above the diagonal
    • AUC depends on score ranking, so a monotonic score transformation can preserve ranking while changing probability values
    • AUC uses only one fixed classification threshold
  8. What does partial AUC let an evaluator emphasize?
    • Only the model's training loss
    • Probability calibration across all score values
    • The number of labels in a multiclass model
    • Performance in a selected false-positive-rate region, such as the low-FPR range