CL4SRec - 리뷰

Contrastive Learning for Sequential Recommendation(2022, ICDE)

논문

• https://www.ijcai.org/proceedings/2019/600

코드

• (참고 코드) https://github.com/RuihongQiu/DuoRec/blob/master/recbole/model/sequential_recommender/cl4srec.py

참고자료

• https://youtu.be/eZRg3QeeiJw?si=1AjiUqwfpii-odRh

문제의식

• Sequential RS suffer from the data sparsity

아이디어

• inspired by recent advances of contrastive learning,
  propose a multi-task model
  • traditional next item prediction task
  • contrastive learning framework
• extract more meaningful patterns and encode the representation effectively

모델 구조

• Original sequence $s$ → augmentation $s^{a}$
  • (3가지 방법) item crop, item mask, item reorder
• encoder : share parameters
• Contrastive Loss : Considering mini-batch (N users → 2N $s^a$)