This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Decoupled Entity Representation Learning for Pinterest Ads Ranking
Created by
Haebom
Author
Jie Liu, Yinrui Li, Jiankai Sun, Kungang Li, Han Sun, Sihan Wang, Huasen Wu, Siyuan Gao, Paulo Soares, Nan Li, Zhifang Liu, Haoyang Li, Siping Ji, Ling Leng, Prathibha Deshikachar
Outline
This paper presents a bottom-up-top-down framework for generating user and pin embeddings from diverse data sources to effectively deliver personalized Pins and ads on Pinterest. The bottom-up model is trained on a wide range of data sources featuring diverse signals and utilizes a complex architecture to capture the complex relationships between users and Pins on Pinterest. To achieve scalability, entity embeddings are learned and periodically updated, rather than computed in real time, allowing asynchronous interaction between the bottom-up and top-down models. These embeddings are integrated as input features into multiple top-down tasks, including ad search and ranking models for CTR and CVR prediction. This framework achieves notable performance gains in both offline and online environments across various top-down tasks, and has been deployed in Pinterest's real-world ad ranking system, yielding significant gains in online metrics.
Takeaways, Limitations
•
Takeaways:
◦
A bottom-up-top-down framework leveraging diverse data sources presents the potential for generating user and pin embeddings and improving personalized advertising effectiveness.
◦
A method for building scalable systems through regular updates of entity embeddings is presented.
◦
Empirically demonstrates performance improvements in both offline and online environments.
◦
Verify online metrics improvements through actual system deployment.
•
Limitations:
◦
Lack of detailed description of the architecture of the specific bottom-up model and the data sources used.
◦
Lack of analysis of the specific mechanisms and performance of asynchronous interactions.
◦
Lack of comparative analysis with other personalization systems.
◦
Lack of discussion about long-term system stability and maintenance.
