README.md

Recommender System WITH PyTorch 🟢🟠🔴

PyTorch Implementation of classic Recommender System Models mainly used for self-learing&communication.

checkout for tensorflow branch

corresponding papers 🚀 RS_Papers 📖

Matching

Matrix Factorization

Model	dataset	loss_func	metrics	state
LFM	ml-100k	MSELoss	MSE: 0.9031	🟢
BiasSVD	ml-100k	MSELoss	MSE: 0.8605	🟢
SVD++	ml-100k	MSELoss	MSE: 0.8493	🟢

Factorization Machine

Model	dataset	loss_func	metrics	state
FM	criteo	BCELoss	AUC: 0.6934	🟢
FFM	criteo	BCELoss	AUC: 0.6729	🟢

Sequential based

Model	dataset	loss_func	metrics	state
FPMC	ml-100k	sBPRLoss	Recall@10: 0.0622	🟢
SASRec	ml-100k	BCEWithLogitsLoss	NDCG@10: 0.1801 HR@10: 0.3595	🟢

Knowledge aware

Model	dataset	loss_func	metrics	state
RippleNet	ml-1m	BCELoss	AUC: 0.8838	🟢

Graph embedding

DeepWalk	Node2vec	EGES

Point of Interests

MIND	SDM

CF

Model	dataset	loss_func	metrics	state
NeuralCF	ml-100k	MSELoss	MSE: 0.3322	🟢

Ranking

FM

Model	dataset	loss_func	metrics	state
FNN	criteo	BCELoss	AUC: 0.6787	🟢
DeepFM	criteo	BCELoss	AUC: 0.6854	🟢
NFM	criteo	BCELoss	AUC: 0.6705	🟢
AFM	criteo	BCELoss	AUC: 0.6572	🟢

LR

GBDT+LR

DNN

Model	dataset	loss_func	metrics	state
Deep Crossing	criteo	BCELoss	AUC: 0.7210	🟢
PNN	criteo	BCELoss	AUC: 0.6360	🟢
Wide&Deep	criteo	BCELoss	AUC: 0.7074	🟢
DCN	criteo	BCELoss	AUC: 0.7335	🟢
DIN	amazon book	BCELoss	AUC: 0.5988	🟢

DIN: It seems that the feature engineering(negative sampling) of paper used for amazon book seems bad. I try hard but the auc of test cannot reach the 0.811 on amazon book.

Multi tasks

Model	dataset	loss_func	metrics	state
MMOE	census-income	BCEWithLogitsLoss	income-AUC: 0.9061 marry-AUC: 0.9637	🟢
ESMM	census-income	BCEWithLogitsLoss	income-ctr-AUC: 0.9242 ctcvr-AUC: 0.9122	🟢