当前位置:网站首页>【Debias】Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in RS(KDD‘21)
【Debias】Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in RS(KDD‘21)
2022-07-25 11:11:00 【chad_lee】
Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System (KDD’21)

图a是我们一般推荐模型的假设,即用户和物品的匹配程度可以反映是否会产生交互。但是事实上应该如图c所示,产生交互不仅仅和匹配程度有关,还和用户和物品本身的偏差有关:物品的流行度、用户是否喜欢流行物品。因此推荐模型应该改造为:
其中user module和item module是一个mlp,推荐模型的输出应该为:
y ^ u i = y ^ k ∗ σ ( y ^ i ) ∗ σ ( y ^ u ) \hat{y}_{u i}=\hat{y}_{k} * \sigma\left(\hat{y}_{i}\right) * \sigma\left(\hat{y}_{u}\right) y^ui=y^k∗σ(y^i)∗σ(y^u)
损失函数的设计为:
L O = ∑ ( u , i ) ∈ D − y u i log ( σ ( y ^ u i ) ) − ( 1 − y u i ) log ( 1 − σ ( y ^ u i ) ) L U = ∑ ( u , i ) ∈ D − y u i log ( σ ( y ^ u ) ) − ( 1 − y u i ) log ( 1 − σ ( y ^ u ) ) L I = ∑ ( u , i ) ∈ D − y u i log ( σ ( y ^ i ) ) − ( 1 − y u i ) log ( 1 − σ ( y ^ i ) ) L = L O + α ∗ L I + β ∗ L U \begin{aligned} L_{O}&=\sum_{(u, i) \in D}-y_{u i} \log \left(\sigma\left(\hat{y}_{u i}\right)\right)-\left(1-y_{u i}\right) \log \left(1-\sigma\left(\hat{y}_{u i}\right)\right)\\ L_{U} &=\sum_{(u, i) \in D}-y_{u i} \log \left(\sigma\left(\hat{y}_{u}\right)\right)-\left(1-y_{u i}\right) \log \left(1-\sigma\left(\hat{y}_{u}\right)\right) \\ L_{I} &=\sum_{(u, i) \in D}-y_{u i} \log \left(\sigma\left(\hat{y}_{i}\right)\right)-\left(1-y_{u i}\right) \log \left(1-\sigma\left(\hat{y}_{i}\right)\right)\\ L&=L_{O}+\alpha * L_{I}+\beta * L_{U} \end{aligned} LOLULIL=(u,i)∈D∑−yuilog(σ(y^ui))−(1−yui)log(1−σ(y^ui))=(u,i)∈D∑−yuilog(σ(y^u))−(1−yui)log(1−σ(y^u))=(u,i)∈D∑−yuilog(σ(y^i))−(1−yui)log(1−σ(y^i))=LO+α∗LI+β∗LU
所以为了消除用户和物品自身的影响,无偏的预测输出应该为:
y ^ k ∗ σ ( y ^ i ) ∗ σ ( y ^ u ) − c ∗ σ ( y ^ i ) ∗ σ ( y ^ u ) \hat{y}_{k} * \sigma\left(\hat{y}_{i}\right) * \sigma\left(\hat{y}_{u}\right)-c * \sigma\left(\hat{y}_{i}\right) * \sigma\left(\hat{y}_{u}\right) y^k∗σ(y^i)∗σ(y^u)−c∗σ(y^i)∗σ(y^u)
边栏推荐
- Start with the development of wechat official account
- php curl post Length Required 错误设置header头
- Maskgae: masked graph modeling meets graph autoencoders
- 【无标题】
- Flinksql client connection Kafka select * from table has no data error, how to solve it?
- The bank's wealth management subsidiary accumulates power to distribute a shares; The rectification of cash management financial products was accelerated
- WIZnet嵌入式以太网技术培训公开课(免费!!!)
- W5500 upload temperature and humidity to onenet platform
- 教你如何通过MCU将S2E配置为UDP的工作模式
- 软件缺陷的管理
猜你喜欢

pycharm连接远程服务器ssh -u 报错:No such file or directory

教你如何通过MCU配置S2E为TCP Server的工作模式

Management of software defects

Reflection reflection

Multi-Label Image Classification(多标签图像分类)

30 sets of Chinese style ppt/ creative ppt templates

brpc源码解析(六)—— 基础类socket详解

【多模态】《TransRec: Learning Transferable Recommendation from Mixture-of-Modality Feedback》 Arxiv‘22

硬件连接服务器 tcp通讯协议 gateway

剑指 Offer 22. 链表中倒数第k个节点
随机推荐
Qin long, a technical expert of Alibaba cloud: a prerequisite for reliability assurance - how to carry out chaos engineering on the cloud
W5500多节点连接
JS中的数组
There is no sound output problem in the headphone jack on the front panel of MSI motherboard [solved]
How to solve the problem that "w5500 chip cannot connect to the server immediately after power failure and restart in tcp_client mode"
The JSP specification requires that an attribute name is preceded by whitespace
Teach you how to configure S2E as the working mode of TCP client through MCU
油猴脚本链接
Functions in JS
JS运算符
11. Reading rumors spread with deep learning
基于W5500实现的考勤系统
JS作用域以及预解析
W5500通过上位机控制实现调节LED灯带的亮度
"Mqtt protocol explanation and Practice (access to onenet)" of wiznet w5500 series training activities
Intelligent information retrieval(智能信息检索综述)
Classification parameter stack of JS common built-in object data types
Maskgae: masked graph modeling meets graph autoencoders
PL/SQL入门,非常详细的笔记
Menu bar + status bar + toolbar ==pyqt5