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
- "Mqtt protocol explanation and Practice (access to onenet)" of wiznet w5500 series training activities
- brpc源码解析(二)—— brpc收到请求的处理过程
- 30 sets of Chinese style ppt/ creative ppt templates
- W5500 upload temperature and humidity to onenet platform
- JS流程控制
- php 一台服务器传图片到另一台上 curl post file_get_contents保存图片
- Miidock Brief
- 教你如何通过MCU将S2E配置为UDP的工作模式
- flink sql client 连接mysql报错异常,如何解决?
- 【多模态】《TransRec: Learning Transferable Recommendation from Mixture-of-Modality Feedback》 Arxiv‘22
猜你喜欢

JaveScript循环

创新突破!亚信科技助力中国移动某省完成核心账务数据库自主可控改造

Maskgae: masked graph modeling meets graph autoencoders

剑指 Offer 22. 链表中倒数第k个节点

brpc源码解析(六)—— 基础类socket详解

How to solve the problem that "w5500 chip cannot connect to the server immediately after power failure and restart in tcp_client mode"

winddows 计划任务执行bat 执行PHP文件 失败的解决办法

MIIdock简述

Intelligent information retrieval(智能信息检索综述)

什么是全局事件总线?
随机推荐
Functions in JS
11. Reading rumors spread with deep learning
Miidock Brief
Teach you how to configure S2E to UDP working mode through MCU
[USB device design] - composite device, dual hid high-speed (64BYTE and 1024byte)
brpc源码解析(一)—— rpc服务添加以及服务器启动主要过程
Eigenvalues and eigenvectors of matrices
LeetCode 50. Pow(x,n)
W5500 multi node connection
toString()与new String()用法区别
30套中国风PPT/创意PPT模板
Intelligent information retrieval(智能信息检索综述)
奉劝那些刚参加工作的学弟学妹们:要想进大厂,这些并发编程知识是你必须要掌握的!完整学习路线!!(建议收藏)
LeetCode第303场周赛(20220724)
JS中的函数
Signal and slot mechanism ==pyqt5
教你如何通过MCU将S2E配置为UDP的工作模式
JS数据类型以及相互转换
Teach you how to configure S2E as the working mode of TCP client through MCU
WIZnet嵌入式以太网技术培训公开课(免费!!!)