MBA

This repository is the implementation of our SIGIR 2023 Paper "Improving Implicit Feedback-Based Recommendation through Multi-Behavior Alignment"

Requirements

• torch == 1.10.1
• scipy == 1.7.1
• numpy
• pandas
• python3

Pre-train

We first need to pre-train two models on multi-behavior data.

Run the shell below to pre-train two MF models on Beibei or Taobao dataset (set --dataset "taobao" to pre-train two models on Taobao dataset).

python main.py --dataset "beibei" --train_method "pre" --model "MF" --lambda0 1e-4 --pretrain_model "MF" --pretrain_early_stop_rounds 20

Run the shell below to pre-train two LightGCN models on Beibei or Taobao dataset (set --dataset "taobao" to pre-train two models on Taobao dataset).

python main.py --dataset "beibei" --train_method "pre" --model "lgn" --lambda0 1e-6 --pretrain_model "lgn" --pretrain_early_stop_rounds 20

MBA-train

After we pre-trained two models on multi-behavior data, we then use them to train our target model with the MBA learning framework.

Run the shell below to train target model.

python main.py --train_method "mba" --dataset ${dataset} --model ${model_name} --pretrain_model ${model_name} --alpha ${alpha} --C_1 ${C_1} --C_2 ${C_2} --lambda0 ${lambda0}  --beta ${beta}

Here we provide a selection of some of the best hyperparameters.

dataset	model_name	$\alpha$	$C_1$	$C_2$	lambda0	$\beta$
beibei	MF	10	100	100	1e-4	0.7(or 0.8)
beibei	lgn	10	10	10	1e-6	1.0
taobao	MF	1	1	1	1e-4	0.8
taobao	lgn	1000	1000	1	1e-6	0.8(or 0.7)

Note: MBD dataset is still under approval as it is the company property.

Contact

If you have any questions for our paper or codes, please send an email to chrisxiangyuan@gmail.com, or raise an issue to me.

Acknowledgement

Our code are developed based on DeCA

Any scientific publications that use our codes should cite our paper as the reference.