Most cross-domain recommenders require samples on target domains or source-target overlaps to carry out domain adaptation. However, in many real-world situations, target domains are lack of such knowledge. Few works discuss this problem, whose essence is domain generalization recommendation. In this paper, we figure out domain generalization recommendation with a clear symbolized definition and propose corresponding models. Moreover, we illustrate its strong connection with zero-shot recommendation, pretrained recommendation and cold-start recommendation, distinguishing it from content-based recommendation. By analyzing its properties, we propose HIRL$^+$ and a series of heuristic methods to solve this problem. We propose hierarchical invariant learning to expel the specific patterns in both domain-level and environment-level, and find the common patterns in generalization space. To make the division of environments flexible, fine-grained and balanced, we put forward a learnable environment assignment method. To improve the robustness against distribution shifts inside domain generalization, we present an adversarial environment refinement method. In addition, we conduct experiments on real-word datasets to verify the effectiveness of our models, and carry out further studies on the domain distance and domain diversity. To benefit the research community and promote this direction, we discuss the future of this field.
In conclusion, the main contributions of this paper can be summarized as follows:
- We give a clear definition of domain generalization recommendation, and demonstrate that it can be considered as a new type of methods for the cold-start problem, pretrained recommendation and zero-shot recommendation.
- We propose HIRL to obtain common and significant patterns on the more fine-grained environment level. We propose learnable environment assignment to make divisions flexible and balanced. We further improve robustness against distribution mismatch by proposing adversarial environment refinement. We put forward HIRL$^+$ with the three components, which is effective on domain generalization recommendation task.
- We conduct experiments on several datasets to compare our model with baselines, and analyze the results. Moreover, we carry out further studies on the domain distance and domain diversity. We release our project at Github.
| Dataset | #User | #Item | #Inter | #Domain |
|---|---|---|---|---|
| CloudTheme-10 | 61,856 | 23,604 | 99,113 | 10 |
| CloudTheme-20 | 107,439 | 44,097 | 172,119 | 20 |
| CloudTheme-50 | 234,444 | 89,386 | 378,641 | 50 |
| KuaiRec-8 | 6,971 | 3,984 | 113,288 | 8 |
| KuaiRec-16 | 7,151 | 8,465 | 189,221 | 16 |
| KuaiRec-all | 7,166 | 8,809 | 216,784 | 20 |
First of all, download our project from github. These files include both codes and datasets.
Create Python 3.9 environment and install the packages that the project requires.
- numpy==1.23.5
- scikit_learn==1.2.0
- torch==1.13.0
You can install the packages with the following command.
pip install -r requirements.txt
Choose a dataset to run (e.g. kuairec_8) and a model (e.g. DNN) to run with the following command.
python run.py DNN dataset/kuairec_8.pickle DNN_kuairec_8 hidden_dim1 16 hidden_dim2 16 batch_size 1024 lr 0.05 epoch 10
You can also use quick_start.py to run the project conveniently.
python quick_start.py
You can also change the hyper-parameters as you want. The necessary hyper-parameters for each model have been recorded in the hyper_parameters.py.
The performance will be saved in result dictionary. The first column is the running time. The second column is the name of models. The third to the fourth columns are metrices NDCG@10, HR@10 in order.
If you find this paper useful, please cite our paper:
@inproceedings{zhang2023hierarchical,
title={Hierarchical Invariant Learning for Domain Generalization Recommendation},
author={Zhang, Zeyu and Gao, Heyang and Yang, Hao and Chen, Xu},
booktitle={Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
pages={3470--3479},
year={2023}
}