/ts

Codebase for paper: "Improving GCN with Transformer layer in social-based items recommendation"

Primary LanguagePython

Python Implementation of Rec-Sys Algorithm

The dataset is a copy of the Epinions and Ciao dataset in the <https://www.cse.msu.edu/~tangjili/datasetcode/truststudy.htm>

Train model

run script below, params are default you can change on main.py to setup model for explore model:

PMF/NeuMF

python main.py --model=<name_model> --dataset=<datasetname>
example:
python main.py --model=PMF --dataset=ciao

GraphREC

python main_graphrec.py --data=data/ciao

GCN

python main_gcn.py --dataset=ciao --epochs=100

GTN

python main_gtn --dataset=ciao

Result

We use 60/20/20 for training, validation, test set

PMF

learning_rate k_dim RMSE MAE
0.001 16 1.05318 0.81547
0.005 16 1.05399 0.81527
0.01 16 1.05903 0.81657
0.05 16 1.06514 0.82039
0.05 8 1.06159 0.82062
0.01 8 1.06556 0.82129
0.005 8 1.05626 0.81623

NeuMF

GraphRec

GTN

Ciao Dataset

Model RMSE MAE
PMF 1.0581 0.8184
NeuMF 1.0439 0.8052
GraphRec 1.0090 0.7834
GCN 1.0605 0.8270
GTN 0.9732 0.7641

Epinions Dataset

Model RMSE MAE
PMF 1.1829 0.9713
NeuMF 1.1411 0.9072
GraphRec 1.1078 0.8524
GCN 1.1680 0.8956
GTN 1.0139 0.8436

Reference:

PMF:

  1. Mnih, A., & Salakhutdinov, R. (2007). Probabilistic matrix factorization. In Advances in neural information processing systems (pp. 1257-1264).
  2. Salakhutdinov, R. Probabilistic matrix factorization in Matlab.

NeuMF:

  1. Neural Collaborative Filtering at WWW'17

GraphRec

  1. Graph Neural Networks for Social Recommendation (WWW'19).

GCN

  1. Graph Convolutional Network

GTN (Graph Convolutional Network add Transformer Encoder layer)

  1. Attention is all you need