xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems【论文记录】

1 摘要

• we propose a novel Compressed Interaction Network (CIN), which aims to generate feature interactions in an explicit fashion and at the vector-wise level.
  我们提出了一种新的压缩交互网络（CIN），旨在以显式方式在矢量级上生成特征交互。

• On one hand, the xDeepFM is able to learn certain bounded-degree feature interactions explicitly; on the other hand, it can learn arbitrary low- and high-order feature interactions implicitly.
  一方面，xDeepFM 能够显式地学习某些边界度特征交互。 另一方面，它可以隐式学习任意的低阶和高阶特征交互。

2 介绍

• Factorization Machines (FM) embed each feature $i$ to a latent factor vector $v_i=[v_{i1},v_{i2},\cdots ,v_{iD}]$，we use the term bit to denote a element (such as $v_{i1}$) in latent vectors
  因子分解机把每个特征 $i$ 都嵌入到隐因子向量 $v_i=[v_{i1},v_{i2},\cdots ,v_{iD}]$，我们用位表示隐向量中的元素（比如 $v_{i1}$）

• there is no theoretical conclusion on what the maximum degree of feature interactions is. DNNs model feature interactions at the bit-wise level, which is different from the traditional FM framework which models feature interactions at the vector-wise level
  没有关于最大特征相互程度的理论结论。 DNN 在按位级别上建模功能交互，这与传统的FM框架不同，后者在矢量级上建模功能交互

• We argue that the CrossNet learns a special type of high-order feature interactions, where each hidden layer in the CrossNet is a scalar multiple of $x_0$. Note that the scalar multiple does not mean ${\mathbf{x}}_k$ is linear with $x_0$. The coefficient ${\alpha }^{i+1}$ is sensitive with ${\mathbf{x}}_0$
  我们认为 CrossNet 学习一种特殊类型的高阶特征交互，其中 CrossNet 中的每个隐藏层都是$x_0$ 的标量倍数。 注意，标量倍数并不意味着 ${\mathbf{x}}_k$ 与 $x_0$ 是线性的。系数 ${\alpha }^{i+1}$ 对 ${\mathbf{x}}_0$ 敏感

3 OUR PROPOSED MODEL

3.1 CIN

• We first apply sum pooling on each feature map of the hidden layer:
  我们首先将求和池应用于隐藏层的每个特征映射：
  $$p_i^k=\sum _{j=1}^D{\mathbf{X}}_{i,j}^k\text{\hspace{1em}(1)}$$ we have a pooling vector $p^k=[p_1^k,p_2^k,\cdots ,p_{H_k}^k]$ with length $H_k$ for the k-th hidden layer
  $H_k$ denotes the number of (embedding) feature vectors in the k-th layer
  $H_k$ 表示第 k 层中的（嵌入）特征向量的数量

• All pooling vectors from hidden layers are concatenated before connected to output units: $p^{+}=[p^1,p^2,\cdots ,p^T]$

3.2 Combination with Implicit Networks$

$$\hat{y}=\sigma \left({\mathbf{w}}_{\text{linear}}^T\mathbf{a}+{\mathbf{w}}_{\text{dnn}}^T{\mathbf{x}}_{\text{dnn}}^k+{\mathbf{w}}_{\text{cin}}^T{\mathbf{p}}^{+}+b\right)\text{\hspace{1em}(2)}$$ CIN 的每一阶特征都输出

4 实验

总结

• we propose a novel network named Compressed Interaction Network (CIN), which aims to learn high-order feature interactions explicitly. CIN has two special virtues: (1) it can learn certain bounded-degree feature interactions effectively; (2) it learns feature interactions at a vector-wise level.
  我们提出了一种新的网络，即压缩交互网络（CIN），其目的是明确地学习高阶特征交互。CIN有两个独特的优点：（1）它可以有效地学习某种有界度的特征交互；（2）在向量水平上学习特征交互。