[ Recommender System ]

12. Wide and Deep Learning for Recommender System

( 참고 : Fastcampus 추천시스템 강의 )

paper : Wide and Deep Learning for Recommender System ( HT Cheng et al., 2016 )

( https://arxiv.org/abs/1606.07792 )

[ Abstract ]

WIDE : 주어진 데이터를 “외운다” ( memorization )

  • cross-product feature transformation

    ( feature간의 interaction 고려 )

  • (단점..?) more feature engineering effort!


DEEP : 일반화 ( generalization )

  • to unseen feature combination

    ( 보지 못한 새로운 조합에 대한 예측 성능 \(\uparrow\) )

  • less feature engineering effort

  • over-generalize


WIDE와 DEEP 둘을 조합하여 좋은 장점을 취한 모델!


1. Introduction

Memorization의 정의

  • Frequent co-occurence of items & features

  • exploit correlation (of historical data)

    \(\rightarrow\) more topical & directly relevant to the items


Generalization의 정의

  • explore new feature combination

    \(\rightarrow\) improve diveristy!


< 기존 모델의 한계 >

  • **1) GLM **
    • ( ex. Logistic Regression )
    • 다양한 feature를 생성하여 학습해야
    • “Memorization”에 focus \(\rightarrow\) overfitting
  • 2) Embedding based Model
    • ( ex. FM, DNN )
    • “Generalization”에 focus \(\rightarrow\) 섬세한 추천 불가


Contributions

  • (1) Wide & Deep Learning Framework 제안 ( NN + LM )
  • (2) Google Play store에 적용 + test
  • (3) 오픈 소스로 제공

figure2

2. Recommender System Overview

figure2

3-1. The WIDE Component

GLM : \(y = w^Tx+b\)

  • \[x = [x_1,..,x_d]\]

    ( 구성 : raw input feature + cross-product feature )

  • \(y\) : 유저의 행동 여부

Cross-product feature

figure2

figure2

3-2. The DEEP Component

Embedding using NN : \(a^{l+1}=f(W^{(1)}a^{(1)}+b^{(1)})\)

figure2

3-2 Joint Training of Wide & Deep Model

figure2


최종적인 output : \(P(Y=1 \mid X)\)

  • app을 다운받을 확류
  • \(P(Y=1 \mid \mathbf{x})=\sigma\left(\mathbf{w}_{w i d e}^{T}[\mathbf{x}, \phi(\mathbf{x})]+\mathbf{w}_{d e e p}^{T} a^{\left(l_{f}\right)}+b\right)\).


Backpropagation은 wide & deep part 모두에게 동시에 이루어진다.

4. System Implementation

Pipeline은 아래와 같다.

figure2

5. Conclusion

  • Wide ( for Memorization ) + Deep ( Generalization )
  • GLM + Embedding NN

Categories:

Updated: