Temel Kavramlar
Precomputed embeddings enhance personalized feeds in real-time, boosting customer engagement and conversions.
Özet
1. Abstract:
- Recommender systems use embeddings for customer actions and items.
- Challenges: user embeddings limit diversity, real-time updates are costly.
- Proposed method: dynamic updates every two minutes using precomputed embeddings.
- Tested at Bol, leading to a 4.9% uplift in conversions.
2. Introduction:
- Bol aims to enhance customer experience with personalized feeds.
- Challenges in personalized feed systems: customer, item, candidate retrieval, and ranking.
- Industry approach and proposed solution for personalized feeds.
3. Four Challenges in Personalized Feed Systems:
- Customer representation challenge: distilling complex behaviors into concise representations.
- Item representation challenge: identifying relevant data for diverse item attributes.
- Candidate retrieval challenge: training customer and item representations, efficient retrieval.
- Ranking challenge: re-ranking retrieved candidates using learning-to-rank algorithms.
4. Our Contributions:
- Overcoming challenges with a user-item framework and dual encoders.
- Addressing drawbacks of user encoding model with multiple embeddings.
- High infrastructure and maintenance costs associated with user embeddings.
5. Related Work:
- Matrix factorization methods pre-deep learning era, deep learning applied in recommender systems.
- YouTube, eBay, Pinterest papers on customer and item representation challenges.
6. Methodology:
- Pfeed involves training multi-vector item embeddings and generating personalized recommendations.
- Representing items with three embeddings: view query, buy query, and target.
- Model architecture: transformer encoder for generating three item embeddings in one model run.
7. Training with Contrastive Learning:
- Training data includes query-target pairs for view-buy and buy-buy relationships.
- Utilizing a Single Input Multi Output (SIMO) embedding model for efficiency.
8. Generating Personalized Feed Recommendations:
- Historical and recent customer interactions used to generate personalized feed recommendations every two minutes.
İstatistikler
사용자 임베딩은 다양성을 제한하고 실시간 업데이트는 비용이 많이 든다.
제안된 방법: 미리 계산된 임베딩을 사용하여 매 두 분마다 동적 업데이트.
Bol에서 테스트되어 고객 참여 및 전환율이 4.9% 향상됨.
Alıntılar
"In personalized recommender systems, embeddings are often used to encode customer actions and items."
"The method enhanced customer engagement and experience, leading to a significant 4.9% uplift in conversions."