This is official code for the WWW 2023 full paper [arXiv]:
ConsRec: Learning Consensus Behind Interactions for Group Recommendation
In this paper, we focus on exploring consensus behind group behavior data. To comprehensively capture the group consensus, we innovatively design three distinct views, including member-level aggregation, item-level tastes, and group-level inherent preferences. Particularly, in the member-level viw, different from existing attentive strategy, we design a novel hypergraph neural network that allows for efficient hypergraph convolutional operations to generate expressive member-level aggregation.
We use two public experimental datasets: Mafengwo and CAMRa2011.
These two datasets' contents are in the data/ folder.
In this paper, we collect a new dataset named Mafengwo-S from Mafengwo to conduct the case study. Particular, in this dataset, we preserve each item's unique semantics, i.e., its location name. It has 11,027 users, 1,236 items, and 1,215 groups.
We would release this dataset soon.
Besides, we release our implementations of group recommendation baselines here.
- Python3
- PyTorch 1.9.1
- scipy 1.6.2
# For Mafengwo
python main.py --dataset=Mafengwo --predictor=MLP --learning_rate=0.0001 --num_negatives=8 --layers=3 --epoch=200
# For CAMRa2011
python main.py --dataset=CAMRa2011 --predictor=DOT --learning_rate=0.001 --num_negatives=2 --layers=2 --epoch=30
For more running options, please refer to main.py
If you make advantage of ConsRec in your research, please cite the following in your manuscript:
@inproceedings{wu2023consrec,
title={ConsRec: Learning Consensus Behind Interactions for Group Recommendation},
author={Wu, Xixi and Xiong, Yun and Zhang, Yao and Jiao, Yizhu and Zhang, Jiawei and Zhu, Yangyong and Philip S. Yu},
booktitle={Proceedings of the ACM Web Conference 2023},
year={2023},
organization={ACM}
}