DeepFM Journal Review

Posted on Disqus:

Journal Review about DeepFM model (pt.1)

DeepFM 모델과 CTR 예측에 관한 논문 하나를 정리해보려고 한다. (사실 리뷰라고 쓰고 정리라 읽는 그런 포스팅이다.)
프로젝트 준비하면서 DeepFM 관련 논문들을 읽어봤지만 사실 무슨 말인지도 모르겠고 과연 코드로 구현할 수 있을지 의문이 들기는 하지만 요약해서 정리해볼까 싶다.

  • Title: A New Approach for Advertising CTR Prediction Based on Deep Neural Network via Attention Mechanism

    (Based on Wang, Q., Liu, F., Xing, S., & Zhao, X. (2018). A New Approach for Advertising CTR Prediction Based on Deep Neural Network via Attention Mechanism. Computational & Mathematical Methods in Medicine, 1–11. https://doi-org.libproxy.library.unt.edu/10.1155/2018/8056541)

  • Introduction:

    • CTR(Click-Through Rate) is critical element to many recommender systems and advertising as estimating the ratio of clicks to impressions of ads, that is to say, how many ads would be displayed to consumers.
    • The Internet market and electronic commerce industry has been rapidly developing, so the online ads with the media are also key issue.
    • As following those development, CTR is an critical indicator to measure the effectiveness of advertising display.
    • Deep learning fields have honored the achievements of NLP, image data processing, so in this article, the authors introduce various methods of CTR prediction based on Deep FM (a part of deep learning) models.

  • Main points (part 3. — This is the main section that I want to emphasize on this post!)

    • Click-through Rate Estimation Based on Deep Neural Network (p.g. 3 to 7)

    To mine highly correlated features, reducing high spareness of features of data, or data dimensional reduction is needed. The complex relationships between various, different types of objects are in data and there is also the similiarity between the same types of objects, so the dimensionality reduction is required.
    In this article, using K-Means clustering algorithm to group/cluster the similar objects into the same cluster for click log data, calculated the distance between user-query-ad (3D vector).

  • I intentionally omitted the formula in this post. (Regarding formula, I will post on another posting. (To be continued)
  • I used this article for personal study and personal project, not the commercial purpose.