Attention, Mask, and Recommendation: A Multi-Level Graph Structure-Aware Method

Environment

To accelerate PPR calculations with GPU, it is recommended to install nx-cugraph. The implementation for PMCF is under the following development environment:

• python=3.10.8
• torch=2.1.2+cu118
• torch-sparse=0.6.18
• networkx=3.3
• numpy=1.26.4
• scipy=1.14.0

Datasets

We utilize three datasets for evaluating PMCF: Yelp, Gowalla, and Amazon. Our evaluation follows the common implicit feedback paradigm. The datasets are divided into training set, validation set and test set by 70:5:25.

Dataset	# Users	# Items	# Interactions	Interaction Density
Yelp	$42,712$	$26,822$	$182,357$	$1.6\times 10^{-4}$
Gowalla	$25,557$	$19,747$	$294,983$	$5.9\times 10^{-4}$
Amazon	$76,469$	$83,761$	$966,680$	$1.5\times 10^{-4}$

Usage

Switch the working directory to methods/PMCF/. The un-specified hyperparameters in the commands are set as default.

• Gowalla

python Main.py --data gowalla --lr 3e-4

• Yelp

python Main.py --data yelp --reg 2e-4 --epoch 150 --gcn_layer 5

• Amazon

python Main.py --data amazon --latdim 72 --head 8 --gcn_layer 3