[논문리뷰] FDSA

Feature-level Deeper Self-Attention Network for Sequential Recommendation(2019, IJCAI)

논문

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

코드

• (RecBole library) https://github.com/RUCAIBox/RecBole/blob/master/recbole/model/sequential_recommender/fdsa.py

참고자료

• 없음

---

문제의식

• Existing Sequential recommendation usually consider the transition patterns between items.
  • ignore the transition patterns between features of items.
  • only the item-level sequences cannot reveal the full sequential patterns.

아이디어

• explicit & implicit feature-level sequences can help extract the full sequential patterns.
  • beneficial for capturing the user’s fine-grained preferences.
  • (ex) a user is more likely to buy shoes after buying clothes
    • the next product’s category is highly related to that of the current product.

모델 구조

1. Instead of using the combined representation of item and its features,
   apply separated self-attention blocks on item sequences and feature sequences, respectively.
   • utilize vanilla attention to a self-attention block to adaptively select essential item's features
2. we combine the contexts at the item-level and the feature-level
   → a fully-connected layer for the recommendation.

• 인풋구성
  • user’s action sequence → a fixed-length sequence
    • item id, category, brand(seller) : one-hot vectors → dense vector representations
    • item's texts (title, description text): topic modeling → Word2vec(word vector)
      • extract five topical keywords vector : apply the Mean Pooling method
• 최종예측
  • a dot product operation
    • $N$ : item embedding matrix
    • $O_{sf}$ : user’s preference for items
      
• 학습
  • binary cross-entropy loss(a negative item sample 활용)
    

(개인의견) 한계점

• 학습시 피처간 관련성 고려 부족
• 추천 설명력 : 성능 포커스