U-BERT - 리뷰

U-BERT: Pre-training User Representations for Improved Recommendation

논문

• https://ojs.aaai.org/index.php/AAAI/article/view/16557

코드

• 공개코드 없음

참고자료

• 없음

---

문제의식

• Building the recommendation system for a particular domain only use the domain
• Domain where the behavior data is insufficient to learn user representations : hurt performance

아이디어

• model user's commenting habits in the domain A and applied them to domain B : better Recommendation
  

학습

• pre-training(content-rich domains) → fine-tuning(target content-insufficient domains)

• Review Encoder : multi-layer Transformer / User Encoder : fusion - attention

• combines 1. user representations, 2. item representations, 3. review interaction information

• [stage 1] pre-training : content-rich domains

  • self-supervision tasks to learn the general user representations
  • (Masked Opinion Token Prediction) Review Encoder
    • we add the additional user representation (to learn the inherent preference of the user)
    • ※ (instead of randomly masking words) we choose the opinion words, shared across domains
  • (Opinion Rating Prediction) User Encoder
    • we use the review-aware user representation (User Encoder output)
    • capture user's general review preference (linking opinions in domains)
  • Loss : the weighted sum of losses of two tasks → multi-task learning

  

• [stage 2] fine-tuning(rating prediction) : target content-insufficient domains

  • encode multiple reviews one-by-one by using the review encoder → row-wise concat
  • Review Co-Matching Layer : measuring their review semantic similarities (Aspect를 고려)

  

성능

• without pre-training (the original BERT’s weights) / pretrain u-bert : 성능 평가