MixRec: Heterogeneous Graph Collaborative Filtering

Environment

The implementation for MixRec is under the following development environment:

• python=3.8.4
• tensorflow=1.14
• numpy=1.22.3
• scipy=1.7.3

Datasets

We utilize three datasets for evaluating MixRec: Beibei, Tmall and IJCAI. We adopt two representative metrics for evaluating the accuracy of top-N item recommendations: Hit Ratio (HR@N) and Normalized Discounted Cumulative Gain (NDCG@N). Following the leave-one-out evaluation strategy, all negative samples are applied to construct the test set with the users’ all positive interactions under the target behavior type.

Datasets	# Users	# Items	# Interactions	Interaction Density
Beibei	21716	7977	282860	0.1633%
Tmall	114503	66706	491870	0.0064%
IJCAI	423423	874328	2926616	0.0008%

Usage

Please unzip the Tmall and IJCAI dataset first. Also you need to create History/ and Models/ directories. Switch the working directory to MixRec/. The command lines to train it on the three datasets are as below. The un-specified hyperparameters in the commands are set as default.

• Beibei

python mixrec_bei.py --data beibei --reg 1 --batch 32 

• Tmall

python mixrec.py --data tmall --ssl_reg 1e-6 --reg 5e-5 --keepRate 0.4 --graphSampleN 20000 --testgraphSampleN 40000

• IJCAI

python mixrec.py --data ijcai --lr 1e-4 --graphSampleN 20000 --testgraphSampleN 40000

Important Arguments

• reg: It is the weight for weight-decay regularization. We tune this hyperparameter from the set {1e-2, 1e-3, 1e-4, 1e-5}.
• ssl_reg and sslGlobal_reg: They are the weights for weight-decay regularization for the node-level and graph-level contrastive objectives. We tune this hyperparameter from the set {1e-4, 1e-5, 1e-6, 1e-7}.
• graphSampleN: This hyperparameter denotes the number of subgraph nodes in the train period. Recommended values are {10000, 15000, 20000, 25000, 30000}.
• testgraphSampleN: This hyperparameter denotes the number of subgraph nodes in the testing. Recommended values are {30000, 35000, 40000, 45000, 50000}.