推荐系统中的常用模型(Recall)
Overview
模型 | 简介 | 论文 |
---|---|---|
Word2Vec | word2vector | [NIPS 2013]Distributed Representations of Words and Phrases and their Compositionality |
DSSM | Deep Structured Semantic Models | [CIKM 2013]Learning Deep Structured Semantic Models for Web Search using Clickthrough Data |
GRU4REC | SR-GRU | [2015]Session-based Recommendations with Recurrent Neural Networks |
Youtube_DNN | Youtube_DNN | [RecSys 2016]Deep Neural Networks for YouTube Recommendations |
SSR | Sequence Semantic Retrieval Model | [SIGIR 2016]Multi-Rate Deep Learning for Temporal Recommendation |
NCF | Neural Collaborative Filtering | [WWW 2017]Neural Collaborative Filtering |
GNN | SR-GNN | [AAAI 2019]Session-based Recommendation with Graph Neural Networks |
Fasttext | fasttext | [EACL 2017]Bag of Tricks for Efficient Text Classification |
Youtube_DNN
输入:
1. 用户观看过的video的embedding向量
2. 用户搜索词的embedding向量
3. 用户的地理位置年龄等静态特征
(这里的embedding向量作者是用word2vec类方法预先生成的)
线下模型训练阶段:
三层ReLU神经网络之后接到softmax层,即建模为为用户推荐下一个感兴趣视频的多分类问题;
输出:所有候选视频集合上的概率分布。
线上预测阶段:
考虑到召回的高性能需求首先通过userId找到相应的用户向量,然后使用KNN类方法找到相似度最高的N条候选结果返回。