#Paper Reading# xDeepFM:Combining Explicit and Implicit Feature Interactions for Recommender Systems

论文题目: xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems
论文地址: https://dl.acm.org/citation.cfm?id=3220023
论文发表于: KDD 2018(CCF A类会议)

论文大体内容:
本文主要介绍了DeepFM模型的变种——xDeepFM(eXtreme Deep Factorization Machine)模型,该模型主要用CIN(Compressed Interaction Network)替换掉DeepFM中的FM,能够学习出高阶、低阶的显式、隐式特征,并在3个真实数据集上取得start-of-art的结果;

1. 本文主要工作点:
①提出xDeepFM模型,同时学习显式和隐式特征;
②提出CIN去学习高阶的显式特征;
③在3个real-world数据集上取得start-of-art的效果;

2. 本文论证了CN(Cross Network)模型不能很好的学习出高阶特征,所以提出了CIN(Compressed Interaction Network)模型。CIN将CN的bit-wise改为vector-wise,并且设计CIN模型主要考虑下面三个方面:
①使用vector-wise;
②能学习到显式高阶特征;
③复杂度不能太高;

3. CIN模型每一层的矩阵递推公式如下,然后该公式可以用CNN思路来理解,就是上面的先计算中间矩阵Z,再对Z做feature map,然后sum pooling,最后得到一个p∈R^(Σ Hi)的向量,可以直接用这个vector来做LR;
#Paper Reading# xDeepFM:Combining Explicit and Implicit Feature Interactions for Recommender Systems_第1张图片
#Paper Reading# xDeepFM:Combining Explicit and Implicit Feature Interactions for Recommender Systems_第2张图片
#Paper Reading# xDeepFM:Combining Explicit and Implicit Feature Interactions for Recommender Systems_第3张图片

4. xDeepFM的整体结构如下,其实xDeepFM=CIN+DNN,CIN模型是本文的主要创新点。CIN学习显式的高阶特征,DNN学习隐式的高阶特征,Linear学习低阶特征;
#Paper Reading# xDeepFM:Combining Explicit and Implicit Feature Interactions for Recommender Systems_第4张图片
#Paper Reading# xDeepFM:Combining Explicit and Implicit Feature Interactions for Recommender Systems_第5张图片

5. 从上面xDeepFM的结构图可以看出,xDeepFM —— (CIN的深度和feature map设置为1) ——> DeepFM ——(去掉DNN)——> FM ;

实验
6. Dataset
#Paper Reading# xDeepFM:Combining Explicit and Implicit Feature Interactions for Recommender Systems_第6张图片

7. Baseline
①LR;
②FM;
③DNN;
④PNN;
⑤WDL;
⑥DCN;
⑦DeepFM;

8. 评测方法
①AUC;
②Logloss

9. 结果
CIN模型效果
#Paper Reading# xDeepFM:Combining Explicit and Implicit Feature Interactions for Recommender Systems_第7张图片
xDeepFM模型效果
#Paper Reading# xDeepFM:Combining Explicit and Implicit Feature Interactions for Recommender Systems_第8张图片

 

以上均为个人见解,因本人水平有限,如发现有所错漏,敬请指出,谢谢!

你可能感兴趣的:(paper,reading,recommender,system,FM,DNN)