2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender

Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender Systems
2021 WSDM

CF的主题是表示学习。学习到的表示应该反映出:

  • 由外在属性所显示的,比如book的作者、类型
  • 由隐式因子所显示的,协同信号

3 问题描述

协作关系感知图,交互图+物品知识图谱

  • 显示关系 R + R^+ R+ 就是物品KG中的关系
  • 隐式关系 R − = { k 1 , k 2 , k 3 , k 4 } R^-=\{k_1,k_2,k_3,k_4\} R={k1,k2,k3,k4} 就是将交互拆解成了几种细粒度的交互,最重要的 k i k_i ki 因素决定了用户采用了特定物品的原因

4 模型

4.1 用户-物品隐式关系建模

2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第1张图片

1 隐式因素分解

2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第2张图片

2 Factor级注意力网络

2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第3张图片

3 注意力 隐式关系-wise 聚合

2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第4张图片

4 解耦正则化

2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第5张图片

4.2 显式关系建模

2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第6张图片

4.3 高阶连通性建模

2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第7张图片

4.4 模型学习

2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第8张图片

5 实验

5.1 实验设置

2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第9张图片
评价指标:

  • Top-N:Recall NDCG {5,10,20}
  • CTR:AUC
  • train:valid:test=6:2:2
  • 取5次实验的平均值

超参数:

  • tanh
  • 交叉熵loss、Adam、Xavier
  • 30 epoch,提前终止策略
  • 对于自己的模型DCF,网格搜索 学习率、解耦项系数、正则化系数
  • 其他模型其他超参数与paper/code设置相同,(有没有网格搜索呢??)

RQ1:模型表现

2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第10张图片
2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第11张图片

RQ2:超参数对模型的影响

2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第12张图片
2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第13张图片

RQ3:设计组件对模型的影响 (消融实验)

2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第14张图片

RQ4:可解释Case Study

2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第15张图片
2021-WSDM-Decomposed Collaborative Filtering: Modeling Explicit and Implicit Factors For Recommender_第16张图片

你可能感兴趣的:(推荐系统,推荐系统)