250113_HW_Labsemimar[Progressive layered extraction (PLE): A novel multi-task learning (MTL) model for personalized recommendations].pptx
- 1. Network Science Lab Dept. of Artificial Intelligence The Catholic University of Korea HYUN-WOO-KIM E-mail: [email protected] Progressive Layered Extraction (PLE): A Novel Multi-Task Learning (MTL) Model for Personalized Recommendations
- 2. Conten ts 2 ⮚ Introduction • Previous work • Negative transfer and Seesaw Phenomenon ⮚ Method • Gate • Shared expert ⮚ Experiment ⮚ Conclusion
- 3. 3 1. Introduction • Personalized recommendation has played a crucial role in online applications. Recommender System(RS) need to incorporate various user feedbacks to model user interests and maximize user engagement and satisfaction. However, user satisfaction is normally hard to tackle directly by a learning algorithm due to the high dimensionality of the problem. Meanwhile, user satisfaction and engagement have many major factors that can be learned directly, e.g. the likelihood of clicking, finishing, sharing, favoriting, and commenting etc. Therefore, there has been an increasing trend to apply Multi-Task Learning (MTL) in RS to model the multiple aspects of user satisfaction or engagement simultaneously
- 4. 4 1. Introduction • In proposed work, there are two problem ○ 1. Negative transfer ■ one task make poor performance of total performance ○ 2. Seesaw phenomenon ■ if one task is optimized, then other task is not optimized like a seesaw • two problem is occurred by these reasons ○ 1. Negative transfer ■ week correlation or no correlation ○ 2. Seesaw phenomenon ■ strong correlation but very complex
- 5. 5 • difference between previous work and PLE is two. ○ gate: extract information that task need. ○ shared expert:sharing common information 1. Introduction
- 6. 6 2. Method • gate select and make combination of output of expert representation
- 7. 7 • there are multi-level extraction networks in PLE to extract higher level shared information. • PLE adopts a progressive separation routing to absorb information from all lower-layer experts, extract higher-level shared knowledge, and separate task-specific parameters gradually.
- 9. 9
- 10. 10
- 11. 11 4. Conclusion • Conclusion ○ In this paper, they propose a novel MTL model called Progressive Layered Extraction (PLE), which separates task-sharing and tasks specific parameters explicitly and introduces an innovative progressive routing manner to avoid the negative transfer and seesaw phenomenon, and achieve more efficient information sharing and joint representation learning.
- 12. 12 5. QnA