Deep Matching：召回策略回顾

Introduction

这篇记录的是推荐系统的召回策略。转载请注明。
reference：WWW18 tutorial

Methods of Representation learning

DSSM

问题：序列的顺序信息丢失，词袋模型不能保证顺序信息

CDSSM，LSTM-DSSM

CNN和RNN都可以维持一些顺序信息。CNN可以维持短期信息，RNN可以维持长期信息。

CNTN

Extension to Representation learning methods

问题：之前的表示学习做文本匹配起来都太过粗糙
解决办法：add-fine-grained signals, include MultGranCNN, U-RAE, MV-LSTM等类似
MV-LSTM

U-RAE

Comparision

Methods of Matching function learning

ARC-II

Problem:
Word level exact matching signals are lost. 2-D matching matrix is constructed based on the embedding of the words in two N-grams.

Matching Pyramid

Inspired by image recognition
Basic matching signals: word-level matching matrix
Matching function:2D-convolution + MLP

Positions information of words is kept

Matching-SRNN

Recursive matching structure

K-NRM

kernel pooling as matching function

Conv-KNRM

Decomposable Attention Model for Matching

Some modern recommendation system architectures

Matrix Factorization as a Neural Network

Deep Matrix Factorization

AutoRec

Deep Collaborative Filtering via Marginalized DAE

Neural Collaborate filtering

NeuMF

NNCF

TransRec

Latent Relation Metric Learning

Wide & Deep

太过熟悉，不提了

Deep Crossing

NFM

太过熟悉，不提了

AFM

太过熟悉，不提了

TreeBased Model

GB-CENT

Deep Embedding Forest

Tree-enhanced Embedding Model

Short Summary

个人觉得其实那么多算法无非就干了两件事情，学习representation和学习matching function。路漫漫其修远兮，还是要多看代码多实现。