SlideShare a Scribd company logo
Interpolative Distillation for Unifying Biased
and Debiased Recommendation
SIGIR’22, Sihao Ding(USTC) et al.
POSTECH DI Lab
Presenter: Changsoo Kwak
2022.5.24
1
Motivation
2
▪ Most recommender system’s test set for evaluate
▪ Normal biased test set(𝐷𝑏)
▪ Debiased test set (𝐷𝑑)
[1] Self-supervised Graph Learning for Recommendation, Jiancan Wu(USTC) et al, SIGIR’21
Existing models didn’t perform well on both test set
Biased or Unbiased model
Only reflect part of whole picture
Intuitive solution?
3
▪ Unifying 𝐷𝑏, 𝐷𝑑
▪ Usually 𝐷𝑏 ≫ |𝐷𝑑|
▪ Train two models for 𝐷𝑏, 𝐷𝑑 respectively, and ensemble
▪ Unclear that each models are strong/weak at which type of users/items
▪ Existing ensemble strategies are not tailored for win-win recommendation scenario
▪ Possible solution?
▪ Distillation!
▪ Aggregate two models at the level of user-item pair
Determine coefficient automatically for distillation
Proposed model(InterD)
4
Environment 𝐸 ∈ {𝑒𝑏, 𝑒𝑑}
Probability of environment given user-item pair
Existing models only consider one environment
- Only achieve good performance on one of 𝐷𝑏 or 𝐷𝑑
Predicted rating with given environment assumption
Let student model learns predicted ratings generated by
fine-grained weighted sum of prediction of pre-trained
models, considering environment
Proposed model(InterD)
5
𝑓𝑏, 𝑓𝑑: Pre-trained biased/unbiased model
▪ Estimate 𝑃(𝑅|𝑈, 𝐼, 𝐸)
▪ Directly use prediction of 𝑓𝑏, 𝑓𝑑
▪ Estimate 𝑃 𝐸 𝑈, 𝐼
𝑤𝑏 =
𝐿𝑏(𝑟𝑏, 𝑟)𝛾1
𝐿𝑏(𝑟𝑏, 𝑟)𝛾1+𝐿𝑑(𝑟𝑑, 𝑟)𝛾1
, 𝑤𝑑 =
𝐿𝑑(𝑟𝑑, 𝑟)𝛾1
𝐿𝑏(𝑟𝑏, 𝑟)𝛾1+𝐿𝑑(𝑟𝑑, 𝑟)𝛾1
𝐿𝑏: MSE, 𝐿𝑑: IPS weighted MSE, 𝛾1: Negative hyperparameter
𝑃 𝑅 𝑈, 𝐼 =
𝐸
𝑃 𝑅 𝑈, 𝐼, 𝐸 𝑃 𝐸 𝑈, 𝐼 = 𝑟∗ = 𝑤𝑏 𝑟𝑏 + 𝑤𝑑𝑟𝑑
▪ Training student model
Distillation loss 𝐿𝑂 =
1
|𝐷𝑏| + |𝐷𝑑|
(𝑢,𝑖,𝑟)∈𝐷𝑏∪𝐷𝑑
𝐿(𝑟, 𝑟∗ )
Proposed model(InterD)
6
▪ Incorporate unobserved data 𝐷𝑛 = 𝑈 × 𝐼 − 𝐷𝑏 ∪ 𝐷𝑑
𝑤𝑏
′
=
𝐿𝑏(𝑟𝑏, 𝑟)𝛾2
𝐿𝑏(𝑟𝑏, 𝑟)𝛾2+𝐿𝑑(𝑟𝑑, 𝑟)𝛾2
, 𝑤𝑑
′
=
𝐿𝑑(𝑟𝑑, 𝑟)𝛾2
𝐿𝑏(𝑟𝑏, 𝑟)𝛾2+𝐿𝑑(𝑟𝑑, 𝑟)𝛾2
𝑟∗
′
= 𝑤𝑏
′
𝑟𝑏 + 𝑤𝑑
′
𝑟𝑑
Imputation distillation loss 𝐿𝑁 =
1
|𝐷𝑛|
(𝑢,𝑖)∈𝐷𝑛
𝐿(𝑟, 𝑟∗
′)
Student model learn more from closer teacher over unobserved data
Experiments
7
Experiments
8

More Related Content

Similar to Review: [SIGIR'22]Interpolative Distillation for Unifying Biased and Debiased Recommendation

NeurIPS22.pptx
NeurIPS22.pptxNeurIPS22.pptx
NeurIPS22.pptx
Julián Tachella
 
Challenging Common Assumptions in the Unsupervised Learning of Disentangled R...
Challenging Common Assumptions in the Unsupervised Learning of Disentangled R...Challenging Common Assumptions in the Unsupervised Learning of Disentangled R...
Challenging Common Assumptions in the Unsupervised Learning of Disentangled R...
Sangwoo Mo
 
Aaa ped-14-Ensemble Learning: About Ensemble Learning
Aaa ped-14-Ensemble Learning: About Ensemble LearningAaa ped-14-Ensemble Learning: About Ensemble Learning
Aaa ped-14-Ensemble Learning: About Ensemble Learning
AminaRepo
 
Lec05.pptx
Lec05.pptxLec05.pptx
Lec05.pptx
HassanAhmad442087
 
DCWP_CVPR2023.pptx
DCWP_CVPR2023.pptxDCWP_CVPR2023.pptx
DCWP_CVPR2023.pptx
건영 박
 
Ensemble methods
Ensemble methods Ensemble methods
Ensemble methods
zekeLabs Technologies
 
adversarial robustness lecture
adversarial robustness lectureadversarial robustness lecture
adversarial robustness lecture
MuhammadAhmedShah2
 
Distributional RL via Moment Matching
Distributional RL via Moment MatchingDistributional RL via Moment Matching
Distributional RL via Moment Matching
taeseon ryu
 
Deep learning paper review ppt sourece -Direct clr
Deep learning paper review ppt sourece -Direct clr Deep learning paper review ppt sourece -Direct clr
Deep learning paper review ppt sourece -Direct clr
taeseon ryu
 
Machine learning - session 3
Machine learning - session 3Machine learning - session 3
Machine learning - session 3
Luis Borbon
 
Conistency of random forests
Conistency of random forestsConistency of random forests
Conistency of random forests
Hoang Nguyen
 
ddpg seminar
ddpg seminarddpg seminar
ddpg seminar
민재 정
 
Week 13 Feature Selection Computer Vision Bagian 2
Week 13 Feature Selection Computer Vision Bagian 2Week 13 Feature Selection Computer Vision Bagian 2
Week 13 Feature Selection Computer Vision Bagian 2
khairulhuda242
 
GTC 2021: Counterfactual Learning to Rank in E-commerce
GTC 2021: Counterfactual Learning to Rank in E-commerceGTC 2021: Counterfactual Learning to Rank in E-commerce
GTC 2021: Counterfactual Learning to Rank in E-commerce
GrubhubTech
 
Learning a nonlinear embedding by preserving class neibourhood structure 최종
Learning a nonlinear embedding by preserving class neibourhood structure   최종Learning a nonlinear embedding by preserving class neibourhood structure   최종
Learning a nonlinear embedding by preserving class neibourhood structure 최종
WooSung Choi
 
I2b2 2008
I2b2 2008I2b2 2008
Basic Concepts of Standard Experimental Designs ( Statistics )
Basic Concepts of Standard Experimental Designs ( Statistics )Basic Concepts of Standard Experimental Designs ( Statistics )
Basic Concepts of Standard Experimental Designs ( Statistics )
Hasnat Israq
 
Multi PPT - Agent Actor-Critic for Mixed Cooperative-Competitive Environments
Multi PPT - Agent Actor-Critic for Mixed Cooperative-Competitive EnvironmentsMulti PPT - Agent Actor-Critic for Mixed Cooperative-Competitive Environments
Multi PPT - Agent Actor-Critic for Mixed Cooperative-Competitive Environments
Jisang Yoon
 
MACHINE LEARNING.pptx
MACHINE LEARNING.pptxMACHINE LEARNING.pptx
MACHINE LEARNING.pptx
SOURAVGHOSH623569
 
BaggingBoosting.pdf
BaggingBoosting.pdfBaggingBoosting.pdf
BaggingBoosting.pdf
DynamicPitch
 

Similar to Review: [SIGIR'22]Interpolative Distillation for Unifying Biased and Debiased Recommendation (20)

NeurIPS22.pptx
NeurIPS22.pptxNeurIPS22.pptx
NeurIPS22.pptx
 
Challenging Common Assumptions in the Unsupervised Learning of Disentangled R...
Challenging Common Assumptions in the Unsupervised Learning of Disentangled R...Challenging Common Assumptions in the Unsupervised Learning of Disentangled R...
Challenging Common Assumptions in the Unsupervised Learning of Disentangled R...
 
Aaa ped-14-Ensemble Learning: About Ensemble Learning
Aaa ped-14-Ensemble Learning: About Ensemble LearningAaa ped-14-Ensemble Learning: About Ensemble Learning
Aaa ped-14-Ensemble Learning: About Ensemble Learning
 
Lec05.pptx
Lec05.pptxLec05.pptx
Lec05.pptx
 
DCWP_CVPR2023.pptx
DCWP_CVPR2023.pptxDCWP_CVPR2023.pptx
DCWP_CVPR2023.pptx
 
Ensemble methods
Ensemble methods Ensemble methods
Ensemble methods
 
adversarial robustness lecture
adversarial robustness lectureadversarial robustness lecture
adversarial robustness lecture
 
Distributional RL via Moment Matching
Distributional RL via Moment MatchingDistributional RL via Moment Matching
Distributional RL via Moment Matching
 
Deep learning paper review ppt sourece -Direct clr
Deep learning paper review ppt sourece -Direct clr Deep learning paper review ppt sourece -Direct clr
Deep learning paper review ppt sourece -Direct clr
 
Machine learning - session 3
Machine learning - session 3Machine learning - session 3
Machine learning - session 3
 
Conistency of random forests
Conistency of random forestsConistency of random forests
Conistency of random forests
 
ddpg seminar
ddpg seminarddpg seminar
ddpg seminar
 
Week 13 Feature Selection Computer Vision Bagian 2
Week 13 Feature Selection Computer Vision Bagian 2Week 13 Feature Selection Computer Vision Bagian 2
Week 13 Feature Selection Computer Vision Bagian 2
 
GTC 2021: Counterfactual Learning to Rank in E-commerce
GTC 2021: Counterfactual Learning to Rank in E-commerceGTC 2021: Counterfactual Learning to Rank in E-commerce
GTC 2021: Counterfactual Learning to Rank in E-commerce
 
Learning a nonlinear embedding by preserving class neibourhood structure 최종
Learning a nonlinear embedding by preserving class neibourhood structure   최종Learning a nonlinear embedding by preserving class neibourhood structure   최종
Learning a nonlinear embedding by preserving class neibourhood structure 최종
 
I2b2 2008
I2b2 2008I2b2 2008
I2b2 2008
 
Basic Concepts of Standard Experimental Designs ( Statistics )
Basic Concepts of Standard Experimental Designs ( Statistics )Basic Concepts of Standard Experimental Designs ( Statistics )
Basic Concepts of Standard Experimental Designs ( Statistics )
 
Multi PPT - Agent Actor-Critic for Mixed Cooperative-Competitive Environments
Multi PPT - Agent Actor-Critic for Mixed Cooperative-Competitive EnvironmentsMulti PPT - Agent Actor-Critic for Mixed Cooperative-Competitive Environments
Multi PPT - Agent Actor-Critic for Mixed Cooperative-Competitive Environments
 
MACHINE LEARNING.pptx
MACHINE LEARNING.pptxMACHINE LEARNING.pptx
MACHINE LEARNING.pptx
 
BaggingBoosting.pdf
BaggingBoosting.pdfBaggingBoosting.pdf
BaggingBoosting.pdf
 

Recently uploaded

Nordic Marketo Engage User Group_June 13_ 2024.pptx
Nordic Marketo Engage User Group_June 13_ 2024.pptxNordic Marketo Engage User Group_June 13_ 2024.pptx
Nordic Marketo Engage User Group_June 13_ 2024.pptx
MichaelKnudsen27
 
Building Production Ready Search Pipelines with Spark and Milvus
Building Production Ready Search Pipelines with Spark and MilvusBuilding Production Ready Search Pipelines with Spark and Milvus
Building Production Ready Search Pipelines with Spark and Milvus
Zilliz
 
Recommendation System using RAG Architecture
Recommendation System using RAG ArchitectureRecommendation System using RAG Architecture
Recommendation System using RAG Architecture
fredae14
 
Energy Efficient Video Encoding for Cloud and Edge Computing Instances
Energy Efficient Video Encoding for Cloud and Edge Computing InstancesEnergy Efficient Video Encoding for Cloud and Edge Computing Instances
Energy Efficient Video Encoding for Cloud and Edge Computing Instances
Alpen-Adria-Universität
 
HCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAU
HCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAUHCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAU
HCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAU
panagenda
 
System Design Case Study: Building a Scalable E-Commerce Platform - Hiike
System Design Case Study: Building a Scalable E-Commerce Platform - HiikeSystem Design Case Study: Building a Scalable E-Commerce Platform - Hiike
System Design Case Study: Building a Scalable E-Commerce Platform - Hiike
Hiike
 
Ocean lotus Threat actors project by John Sitima 2024 (1).pptx
Ocean lotus Threat actors project by John Sitima 2024 (1).pptxOcean lotus Threat actors project by John Sitima 2024 (1).pptx
Ocean lotus Threat actors project by John Sitima 2024 (1).pptx
SitimaJohn
 
Skybuffer AI: Advanced Conversational and Generative AI Solution on SAP Busin...
Skybuffer AI: Advanced Conversational and Generative AI Solution on SAP Busin...Skybuffer AI: Advanced Conversational and Generative AI Solution on SAP Busin...
Skybuffer AI: Advanced Conversational and Generative AI Solution on SAP Busin...
Tatiana Kojar
 
Overcoming the PLG Trap: Lessons from Canva's Head of Sales & Head of EMEA Da...
Overcoming the PLG Trap: Lessons from Canva's Head of Sales & Head of EMEA Da...Overcoming the PLG Trap: Lessons from Canva's Head of Sales & Head of EMEA Da...
Overcoming the PLG Trap: Lessons from Canva's Head of Sales & Head of EMEA Da...
saastr
 
UI5 Controls simplified - UI5con2024 presentation
UI5 Controls simplified - UI5con2024 presentationUI5 Controls simplified - UI5con2024 presentation
UI5 Controls simplified - UI5con2024 presentation
Wouter Lemaire
 
Trusted Execution Environment for Decentralized Process Mining
Trusted Execution Environment for Decentralized Process MiningTrusted Execution Environment for Decentralized Process Mining
Trusted Execution Environment for Decentralized Process Mining
LucaBarbaro3
 
Serial Arm Control in Real Time Presentation
Serial Arm Control in Real Time PresentationSerial Arm Control in Real Time Presentation
Serial Arm Control in Real Time Presentation
tolgahangng
 
Your One-Stop Shop for Python Success: Top 10 US Python Development Providers
Your One-Stop Shop for Python Success: Top 10 US Python Development ProvidersYour One-Stop Shop for Python Success: Top 10 US Python Development Providers
Your One-Stop Shop for Python Success: Top 10 US Python Development Providers
akankshawande
 
Deep Dive: Getting Funded with Jason Jason Lemkin Founder & CEO @ SaaStr
Deep Dive: Getting Funded with Jason Jason Lemkin Founder & CEO @ SaaStrDeep Dive: Getting Funded with Jason Jason Lemkin Founder & CEO @ SaaStr
Deep Dive: Getting Funded with Jason Jason Lemkin Founder & CEO @ SaaStr
saastr
 
Azure API Management to expose backend services securely
Azure API Management to expose backend services securelyAzure API Management to expose backend services securely
Azure API Management to expose backend services securely
Dinusha Kumarasiri
 
Generating privacy-protected synthetic data using Secludy and Milvus
Generating privacy-protected synthetic data using Secludy and MilvusGenerating privacy-protected synthetic data using Secludy and Milvus
Generating privacy-protected synthetic data using Secludy and Milvus
Zilliz
 
Choosing The Best AWS Service For Your Website + API.pptx
Choosing The Best AWS Service For Your Website + API.pptxChoosing The Best AWS Service For Your Website + API.pptx
Choosing The Best AWS Service For Your Website + API.pptx
Brandon Minnick, MBA
 
Monitoring and Managing Anomaly Detection on OpenShift.pdf
Monitoring and Managing Anomaly Detection on OpenShift.pdfMonitoring and Managing Anomaly Detection on OpenShift.pdf
Monitoring and Managing Anomaly Detection on OpenShift.pdf
Tosin Akinosho
 
Best 20 SEO Techniques To Improve Website Visibility In SERP
Best 20 SEO Techniques To Improve Website Visibility In SERPBest 20 SEO Techniques To Improve Website Visibility In SERP
Best 20 SEO Techniques To Improve Website Visibility In SERP
Pixlogix Infotech
 
Presentation of the OECD Artificial Intelligence Review of Germany
Presentation of the OECD Artificial Intelligence Review of GermanyPresentation of the OECD Artificial Intelligence Review of Germany
Presentation of the OECD Artificial Intelligence Review of Germany
innovationoecd
 

Recently uploaded (20)

Nordic Marketo Engage User Group_June 13_ 2024.pptx
Nordic Marketo Engage User Group_June 13_ 2024.pptxNordic Marketo Engage User Group_June 13_ 2024.pptx
Nordic Marketo Engage User Group_June 13_ 2024.pptx
 
Building Production Ready Search Pipelines with Spark and Milvus
Building Production Ready Search Pipelines with Spark and MilvusBuilding Production Ready Search Pipelines with Spark and Milvus
Building Production Ready Search Pipelines with Spark and Milvus
 
Recommendation System using RAG Architecture
Recommendation System using RAG ArchitectureRecommendation System using RAG Architecture
Recommendation System using RAG Architecture
 
Energy Efficient Video Encoding for Cloud and Edge Computing Instances
Energy Efficient Video Encoding for Cloud and Edge Computing InstancesEnergy Efficient Video Encoding for Cloud and Edge Computing Instances
Energy Efficient Video Encoding for Cloud and Edge Computing Instances
 
HCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAU
HCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAUHCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAU
HCL Notes und Domino Lizenzkostenreduzierung in der Welt von DLAU
 
System Design Case Study: Building a Scalable E-Commerce Platform - Hiike
System Design Case Study: Building a Scalable E-Commerce Platform - HiikeSystem Design Case Study: Building a Scalable E-Commerce Platform - Hiike
System Design Case Study: Building a Scalable E-Commerce Platform - Hiike
 
Ocean lotus Threat actors project by John Sitima 2024 (1).pptx
Ocean lotus Threat actors project by John Sitima 2024 (1).pptxOcean lotus Threat actors project by John Sitima 2024 (1).pptx
Ocean lotus Threat actors project by John Sitima 2024 (1).pptx
 
Skybuffer AI: Advanced Conversational and Generative AI Solution on SAP Busin...
Skybuffer AI: Advanced Conversational and Generative AI Solution on SAP Busin...Skybuffer AI: Advanced Conversational and Generative AI Solution on SAP Busin...
Skybuffer AI: Advanced Conversational and Generative AI Solution on SAP Busin...
 
Overcoming the PLG Trap: Lessons from Canva's Head of Sales & Head of EMEA Da...
Overcoming the PLG Trap: Lessons from Canva's Head of Sales & Head of EMEA Da...Overcoming the PLG Trap: Lessons from Canva's Head of Sales & Head of EMEA Da...
Overcoming the PLG Trap: Lessons from Canva's Head of Sales & Head of EMEA Da...
 
UI5 Controls simplified - UI5con2024 presentation
UI5 Controls simplified - UI5con2024 presentationUI5 Controls simplified - UI5con2024 presentation
UI5 Controls simplified - UI5con2024 presentation
 
Trusted Execution Environment for Decentralized Process Mining
Trusted Execution Environment for Decentralized Process MiningTrusted Execution Environment for Decentralized Process Mining
Trusted Execution Environment for Decentralized Process Mining
 
Serial Arm Control in Real Time Presentation
Serial Arm Control in Real Time PresentationSerial Arm Control in Real Time Presentation
Serial Arm Control in Real Time Presentation
 
Your One-Stop Shop for Python Success: Top 10 US Python Development Providers
Your One-Stop Shop for Python Success: Top 10 US Python Development ProvidersYour One-Stop Shop for Python Success: Top 10 US Python Development Providers
Your One-Stop Shop for Python Success: Top 10 US Python Development Providers
 
Deep Dive: Getting Funded with Jason Jason Lemkin Founder & CEO @ SaaStr
Deep Dive: Getting Funded with Jason Jason Lemkin Founder & CEO @ SaaStrDeep Dive: Getting Funded with Jason Jason Lemkin Founder & CEO @ SaaStr
Deep Dive: Getting Funded with Jason Jason Lemkin Founder & CEO @ SaaStr
 
Azure API Management to expose backend services securely
Azure API Management to expose backend services securelyAzure API Management to expose backend services securely
Azure API Management to expose backend services securely
 
Generating privacy-protected synthetic data using Secludy and Milvus
Generating privacy-protected synthetic data using Secludy and MilvusGenerating privacy-protected synthetic data using Secludy and Milvus
Generating privacy-protected synthetic data using Secludy and Milvus
 
Choosing The Best AWS Service For Your Website + API.pptx
Choosing The Best AWS Service For Your Website + API.pptxChoosing The Best AWS Service For Your Website + API.pptx
Choosing The Best AWS Service For Your Website + API.pptx
 
Monitoring and Managing Anomaly Detection on OpenShift.pdf
Monitoring and Managing Anomaly Detection on OpenShift.pdfMonitoring and Managing Anomaly Detection on OpenShift.pdf
Monitoring and Managing Anomaly Detection on OpenShift.pdf
 
Best 20 SEO Techniques To Improve Website Visibility In SERP
Best 20 SEO Techniques To Improve Website Visibility In SERPBest 20 SEO Techniques To Improve Website Visibility In SERP
Best 20 SEO Techniques To Improve Website Visibility In SERP
 
Presentation of the OECD Artificial Intelligence Review of Germany
Presentation of the OECD Artificial Intelligence Review of GermanyPresentation of the OECD Artificial Intelligence Review of Germany
Presentation of the OECD Artificial Intelligence Review of Germany
 

Review: [SIGIR'22]Interpolative Distillation for Unifying Biased and Debiased Recommendation

  • 1. Interpolative Distillation for Unifying Biased and Debiased Recommendation SIGIR’22, Sihao Ding(USTC) et al. POSTECH DI Lab Presenter: Changsoo Kwak 2022.5.24 1
  • 2. Motivation 2 ▪ Most recommender system’s test set for evaluate ▪ Normal biased test set(𝐷𝑏) ▪ Debiased test set (𝐷𝑑) [1] Self-supervised Graph Learning for Recommendation, Jiancan Wu(USTC) et al, SIGIR’21 Existing models didn’t perform well on both test set Biased or Unbiased model Only reflect part of whole picture
  • 3. Intuitive solution? 3 ▪ Unifying 𝐷𝑏, 𝐷𝑑 ▪ Usually 𝐷𝑏 ≫ |𝐷𝑑| ▪ Train two models for 𝐷𝑏, 𝐷𝑑 respectively, and ensemble ▪ Unclear that each models are strong/weak at which type of users/items ▪ Existing ensemble strategies are not tailored for win-win recommendation scenario ▪ Possible solution? ▪ Distillation! ▪ Aggregate two models at the level of user-item pair Determine coefficient automatically for distillation
  • 4. Proposed model(InterD) 4 Environment 𝐸 ∈ {𝑒𝑏, 𝑒𝑑} Probability of environment given user-item pair Existing models only consider one environment - Only achieve good performance on one of 𝐷𝑏 or 𝐷𝑑 Predicted rating with given environment assumption Let student model learns predicted ratings generated by fine-grained weighted sum of prediction of pre-trained models, considering environment
  • 5. Proposed model(InterD) 5 𝑓𝑏, 𝑓𝑑: Pre-trained biased/unbiased model ▪ Estimate 𝑃(𝑅|𝑈, 𝐼, 𝐸) ▪ Directly use prediction of 𝑓𝑏, 𝑓𝑑 ▪ Estimate 𝑃 𝐸 𝑈, 𝐼 𝑤𝑏 = 𝐿𝑏(𝑟𝑏, 𝑟)𝛾1 𝐿𝑏(𝑟𝑏, 𝑟)𝛾1+𝐿𝑑(𝑟𝑑, 𝑟)𝛾1 , 𝑤𝑑 = 𝐿𝑑(𝑟𝑑, 𝑟)𝛾1 𝐿𝑏(𝑟𝑏, 𝑟)𝛾1+𝐿𝑑(𝑟𝑑, 𝑟)𝛾1 𝐿𝑏: MSE, 𝐿𝑑: IPS weighted MSE, 𝛾1: Negative hyperparameter 𝑃 𝑅 𝑈, 𝐼 = 𝐸 𝑃 𝑅 𝑈, 𝐼, 𝐸 𝑃 𝐸 𝑈, 𝐼 = 𝑟∗ = 𝑤𝑏 𝑟𝑏 + 𝑤𝑑𝑟𝑑 ▪ Training student model Distillation loss 𝐿𝑂 = 1 |𝐷𝑏| + |𝐷𝑑| (𝑢,𝑖,𝑟)∈𝐷𝑏∪𝐷𝑑 𝐿(𝑟, 𝑟∗ )
  • 6. Proposed model(InterD) 6 ▪ Incorporate unobserved data 𝐷𝑛 = 𝑈 × 𝐼 − 𝐷𝑏 ∪ 𝐷𝑑 𝑤𝑏 ′ = 𝐿𝑏(𝑟𝑏, 𝑟)𝛾2 𝐿𝑏(𝑟𝑏, 𝑟)𝛾2+𝐿𝑑(𝑟𝑑, 𝑟)𝛾2 , 𝑤𝑑 ′ = 𝐿𝑑(𝑟𝑑, 𝑟)𝛾2 𝐿𝑏(𝑟𝑏, 𝑟)𝛾2+𝐿𝑑(𝑟𝑑, 𝑟)𝛾2 𝑟∗ ′ = 𝑤𝑏 ′ 𝑟𝑏 + 𝑤𝑑 ′ 𝑟𝑑 Imputation distillation loss 𝐿𝑁 = 1 |𝐷𝑛| (𝑢,𝑖)∈𝐷𝑛 𝐿(𝑟, 𝑟∗ ′) Student model learn more from closer teacher over unobserved data

Editor's Notes

  1. RCT: Randomized Control Trial(https://books.google.co.kr/books?id=JUTqDwAAQBAJ&pg=PA244&lpg=PA244&dq=yahoo!r3+randomized+controlled+trial&source=bl&ots=0cagKMc4KG&sig=ACfU3U3oFb-FZsxO3PuYDFYRz6gX9O97tA&hl=ko&sa=X&ved=2ahUKEwj5qp-psev3AhWim1YBHfVgC2QQ6AF6BAgDEAM#v=onepage&q=yahoo!r3%20randomized%20controlled%20trial&f=false)
  2. In other words, the student tends to learn the easier aspects of knowledge since the smaller distance makes it easier to follow the corresponding teacher 학생 입장에서 더 쉬운 쪽(거리가 적은 쪽 teacher)을 따라가기 때문에 curriculum learning으로 볼 수도 있다? Weight 계산에 student prediction이 들어가니까 self-paced learning으,로 볼 수도 있다?