An Empirical Study Towards Prompt-Tuning for Graph Contrastive Pre-Training in Recommendations
Authors: Haoran Yang, Xiangyu Zhao, Yicong Li, Hongxu Chen, Guandong Xu
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our extensive experiments on three benchmark datasets confirm the effectiveness of CPTPP compared to state-of-the-art baselines. |
| Researcher Affiliation | Academia | Haoran Yang University of Technology Sydney haoran.yang-2@student.uts.edu.au Xiangyu Zhao City University of Hong Kong xianzhao@cityu.edu.hk Yicong Li University of Technology Sydney yicong.li@student.uts.edu.au Hongxu Chen University of Queensland hongxu.chen@uq.edu.au Guandong Xu University of Technology Sydney guandong.xu@uts.edu.au |
| Pseudocode | Yes | Algorithm 1: CPTPP algorithm |
| Open Source Code | Yes | The implementation code is available online2 for reproducibility. 2https://github.com/Haoran-Young/CPTPP |
| Open Datasets | Yes | To verify the performance of CPTPP in the recommendation task, we select three popular datasets: Douban [34], Movie Lens-1M [5], and Gowalla [13]. The detailed statistics about the three datasets are listed in Table 1. For each dataset, we randomly select 80% of historical user-item interactions as the training set, and the rest 20% records will serve as the testing set. Appendix A.1 provides the links for downloading these datasets: Douban: https://pan.baidu.com/s/1hr JP6rq#list/path=%2F ML-1M: https://grouplens.org/datasets/movielens/1m/ Gowalla: https://github.com/kuandeng/Light GCN/tree/master/Data/gowalla |
| Dataset Splits | Yes | For each dataset, we randomly select 80% of historical user-item interactions as the training set, and the rest 20% records will serve as the testing set. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers. |
| Experiment Setup | Yes | To ensure reproducibility, we disclose the comprehensive hyperparameter settings for implementing our proposed CPTPP in the source codes and Appendix A.3. Table 3: Summary of hyper-parameter settings of CPTPP. Hidden dimension size d 64 64 64 Pre-train epoch 10 10 10 Prompt-tune epoch 100 100 100 Batch size 512 512 2048 Learning rate 0.003 0.001 0.001 Regularizer weight λ 0.0001 0.0001 0.0001 Number of GNN layers 2 2 2 Dropout rate 0.1 0.1 0.1 Temperature parameter τ 0.2 0.2 0.2 Prompt size p {8, 16, 32, 64, 128, 256} |