lawRossi / recommendation

recommendation models

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Licence Python

简介

跟踪推荐系统研究,用pytorch实现部分推荐模型并进行实验。

已实现模型

召回模型

模型名称 说明 参考文献
Youtube-Net Youtube提出的经典的召回模型 Covington, Paul, Jay Adams, and Emre Sargin. "Deep neural networks for youtube recommendations." Proceedings of the 10th ACM conference on recommender systems. 2016.
GES 阿里提出的Graph Embedding with Side-information模型 Wang, Jizhe, et al. "Billion-scale commodity embedding for e-commerce recommendation in alibaba." Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2018.
EGES 阿里提出的Enhanced Graph Embedding with Side-information模型 Wang, Jizhe, et al. "Billion-scale commodity embedding for e-commerce recommendation in alibaba." Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2018.

排序模型

模型名称 description reference
NRMS 微软研究研究提出的新闻推荐模型 Wu, Chuhan, et al. "Neural news recommendation with multi-head self-attention." Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.
BST 阿里提出的Behavior Sequence Transformer模型 Chen, Qiwei, et al. "Behavior sequence transformer for e-commerce recommendation in alibaba." Proceedings of the 1st International Workshop on Deep Learning Practice for High-Dimensional Sparse Data. 2019.

TODO

  1. 实现DIN模型
  2. 拿实验数据对比各个模型效果

模型效果

  1. 新闻推荐

实验设置

  1. 数据集 使用微软发布的新闻推荐数据集

  2. 模型参数

最大历史点击数:30
标题最大长度(词数): 25
dropout: 0.2
embedding 维度300
注意力头:20
addative attention隐含层维度:200

模型名称 dev-set auc 备注
NRMS 0.671 论文提出的原始模型
NRMS-cosine 0.684 基于原始模型修改了相似度计算和loss计算

About

recommendation models

License:MIT License


Languages

Language:Python 100.0%