Skip to content

Latest commit

 

History

History
63 lines (36 loc) · 5.04 KB

candidate_matching.md

File metadata and controls

63 lines (36 loc) · 5.04 KB
Raw

Candidate Item Matching

Two-Tower Networks

  • [EBR][Facebook] Embedding-based Retrieval in Facebook Search, by Jui-Ting Huang, Ashish Sharma, Shuying Sun, Li Xia, David Zhang, Philip Pronin, Janani Padmanabhan, Giuseppe Ottaviano, Linjun Yang. KDD 2020.

    本文介绍了Facebook的语义检索模型的工程实践经验，包括：1) 随机负采样比曝光未点击样本好 2) 采用triplet loss 3) ANN引擎加入了条件binary matching的过滤 4) retrieval阶段变了，ranking阶段也要跟着优化，考虑加入embedding做特征(query-doc的cosine效果好)和利用人工标注数据finetune embeddings 5) Online hard mining有效(选取相似度最大的negative)；100:1的easy:hard mixing效果最好 6) 双阶段retrieval效果也有提升

Path-based Networks

  • [PDN][Alibaba] SIGIR 2021.

    本文利用U2I和I2I的三部图来进行分数计算，user和target的分数分为U2I(用户与交互过的item之间)和I2I(交互过的item与target item)的乘积之和。分别由TriggerNet和SimNet两个网络进行学习。

Multi-Interest Recommendation

  • [PinnerSage][Pinterest] PinnerSage: Multi-Modal User Embedding Framework for Recommendations at Pinterest, KDD 2020.

    本文介绍了Pinterest的多兴趣召回方案，先采用PinSage学习item embedding，作为输入采用Ward聚类方法对历史用户序列(90天)进行聚类，Ward类似hierarchical clustering，聚类数是不定的。然后获取聚类的的中心作为兴趣表征，同时根据类别的大小及行为时间定义了类别的importance，召回时根据importance选取多个兴趣进行召回。同时，从系统上考虑了batch inference和online update来获取实时用户行为更新聚类中心。与单兴趣(single embedding)召回相比，线上效果显著。

  • [ComiRec][Alibaba] Controllable Multi-Interest Framework for Recommendation, by Yukuo Cen, Jianwei Zhang, Xu Zou, Chang Zhou, Hongxia Yang, Jie Tang. KDD 2020.

  • [MIND][Alibaba] Multi-Interest Network with Dynamic Routing for Recommendation at Tmall, by Chao Li, Zhiyuan Liu, Mengmeng Wu, Yuchi Xu, Pipei Huang, Huan Zhao, Guoliang Kang, Qiwei Chen, Wei Li, Dik Lun Lee. CIKM 2019

dfd

Sequential Recommendation

  1. [CIKM'19] BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer, by Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, Peng Jiang. [Alibaba]

Graph-based Recommendation

  1. [Alibaba] Beyond Keywords and Relevance: A Personalized Ad Retrieval Framework in E-Commerce Sponsored Search, by Su Yan, Wei Lin, Tianshu Wu, Daorui Xiao, Xu Zheng, Bo Wu, Kaipeng Liu. WWW 2018.

    本文提出了新的一种ad召回框架，传统的广告召回都靠bid keywords, 但是广告主并不能列举出所有的ad keywords，导致有些高度相关的ad却无法召回的情况。现有的召回框架一般用kewords和relevance的方法。相反，本文提出采用signal nodes--key nodes--ad nodes的异构网络的挖掘，包括query到key node的rewriting边和key node到ad的ad-selecting边。最后对这些边建立index，用于快速的召回查询。与纯文本召回相比，该方法引入了用户历史点击行为，更偏向CTR的优化。[Read more...]

End-to-End Retrieval

  1. [Alibaba] Joint Optimization of Tree-based Index and Deep Model for Recommender Systems, by Han Zhu, Daqing Chang, Ziru Xu, Pengye Zhang, Xiang Li, Jie He, Han Li, Jian Xu, Kun Gai. NeurIPS 2019.

  2. [TDM][Alibaba] Han Zhu, Xiang Li, Pengye Zhang, Guozheng Li, Jie He, Han Li, Kun Gai. Learning Tree-based Deep Model for Recommender Systems, KDD 2018.

Lookalike Model

  1. [Tencent][RALM] Real-time Attention Based Look-alike Model for Recommender System, by Yudan Liu, Kaikai Ge, Xu Zhang, Leyu Lin. KDD 2019.

Cross-Domain Recommendation

  1. [KDD'18] Learning and Transferring IDs Representation in E-commerce, by Kui Zhao, Yuechuan Li, Zhaoqian Shuai, Cheng Yang. [Alibaba]

Interactive Recommendation

  1. [Google][KDD'18] Q&R: A Two-Stage Approach Toward Interactive Recommendation, by Konstantina Christakopoulou, Alex Beutel, Rui Li, Sagar Jain, Ed H. Chi.