No Story, No Ecstasy

Ad Segments beyond Look-alike Audience Model 본문

Data Science Series

Ad Segments beyond Look-alike Audience Model

heave_17 2021. 9. 7. 00:59

https://towardsdatascience.com/data-science-for-ad-segments-moving-beyond-look-alike-modeling-fc0b97ed8bb6

 

Data Science for Ad Segments : Moving Beyond Look-Alike Modeling

Segment Classification, Conversion Predictions, and Uplift Modeling

towardsdatascience.com

 

Look Alike Segment?

  - 주어진 유저 셋과 가장 유사하게 '보이는' 청중을 찾는 기법

  - 보통 2개의 정보를 필요로 한다

    1. Seed Set

    2. Segment Size

 

Limitation?

  - PU learning과 같은 semi-supervised ML -> seed set의 정보만으로 모델링한다.

  - 당연히 positive set에 bias가 생길 수 밖에 없다. -> Training의 정확도는 높으나 실제로는 무쓸모일 수도..

  - 초기에 데이터가 너무 없어서 케이스를 모으기 위해 사용될 수는 있다.

 

Alternative #1: Classification

  - Positive vs Negative data를 모두 가지고 모델링을 수행하기에 bias가 줄어든다.

 

Alternative #2: Conversion Predictions

  - 향후 Positive set으로 변화될 후보 데이터를 예측하는데 사용될 수 있다.

 

Alternative #3: Uplift Predictions

  - Uplift (증가, 희망) model is a prescriptive technique which predicts how each customer is likely to respond to a marketing action. 

 

 

It's time to stop wasting marketing spend and start using uplift models - Marketing Tech News

Uplift modeling is the next frontier for data-driven marketers as they cut right to the core: how will each customer respond to a marketing campaign?

marketingtechnews.net

  - 특정 광고가 효과적일 것 같은 청중을 타겟팅할 때 효과적일 수 있음