giangnguyen2412 / InterpretableMLBook-Vietnamese

Bản dịch của cuốn "Interpretable Machine Learning: A Guide for Making Black Box Models Explainable" sang tiếng Việt
110 stars 44 forks source link

Chapter 5.8 #12

Open tungnguyen1234 opened 4 years ago

tungnguyen1234 commented 4 years ago

Chào mọi người em có dịch qua một số từ trong phần 5.8 như sau. Bốn từ đầu được lấy từ đoạn này:

Like its predecessor, the anchors approach deploys a perturbation-based strategy to generate local explanations for predictions of black-box machine learning models. However, instead of surrogate models used by LIME, the resulting explanations are expressed as easy-to-understand IF-THEN rules, called anchors. These rules are reusable since they are scoped: anchors include the notion of coverage, stating precisely to which other, possibly unseen, instances they apply.

Các từ còn lại em dịch như sau:

Em cảm ơn!

giangnguyen2412 commented 4 years ago

Những từ còn lại anh đồng ý!

tungnguyen1234 commented 4 years ago

Dạ em có thêm câu hỏi là có một phần dịch như sau:

An anchor A is formally defined as follows:

image

Ở đây có hai từ "queried" hay "local fidelity" không hề có trong bài báo gốc. Tứ "queried" nó không làm tăng ý nghĩa và trong bài báo gốc chỉ refer $f$ là black-box model, cho nên em nghĩ bỏ từ này luôn cũng OK.

Từ focal fidelity thì em tạm dịch là trung thành lân cận nhưng vẫn rối nghĩa. Vì tác gỉa chỉ dùng từ này đúng 1 lần, và phần sau đã giải thích công thức như sau:

image

Thì em nghĩ bỏ luôn toàn bộ câu có "local fidelity" cũng không làm mất nghĩa.

Mọi người thấy vậy OK không ạ?

giangnguyen2412 commented 4 years ago

a đang định fidelity là tính (độ) nhất quán (trong context của local surrogate models). queried trong context đang dịch a nghĩ mình có thể dịch là sử dụng.