Curriculum Conditioned Diffusion for Multimodal Recommendation
Authors: Yimeng Yang, Haokai Ma, Lei Meng, Shuo Xu, Ruobing Xie, Xiangxu Meng
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on three datasets with four diverse backbones demonstrate the effectiveness and robustness of our CCDRec. The visualization analyses also clarify the underlying mechanism of our DMA in multimodal representation alignment and CNS in curricular negative discovery. |
| Researcher Affiliation | Collaboration | 1School of Software, Shandong University, Jinan, China 2Shandong Research Institute of Industrial Technology, Jinan, China 3Tencent, China EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology in prose and mathematical equations but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The code and the corresponding dataset will be uploaded in the Appendix. |
| Open Datasets | Yes | The code and the corresponding dataset will be uploaded in the Appendix. Following previous works (Zhou 2023), we perform experiments on the Baby, Sports, and Clothing datasets from the Amazon platform. We pre-process the data with a 5-core setting on items and users, as used in (He and Mc Auley 2016), and present the results in Table 1. |
| Dataset Splits | No | The paper mentions pre-processing with a '5-core setting' but does not specify training, validation, or test dataset splits (e.g., percentages or exact counts) needed for reproduction. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU or CPU models, memory specifications, or cloud computing resources used for experiments. |
| Software Dependencies | No | The paper mentions 'sentence-transformers' for obtaining textual features but does not provide specific version numbers for this or any other software libraries or dependencies. |
| Experiment Setup | Yes | Following the classical works (Zhang et al. 2021; Zhou et al. 2023; Zhou and Shen 2023), we set the embedding size of both users and items to 64 for all models. To ensure a fair comparison, we present the results of other methods using two random negative samples. We perform a comprehensive grid search to select the optimal universal hyper-parameters. To be specific, the number of GCN layers is set to 2. We set the loss weight λ at {0.5, 1, 2} and ω at {0.5, 0.7, 0.8, 0.9}. As for the diffusion process, the step t is tuned in {5, 10, 20, 40, 100}. Respectively, the diffused weight µ is chosen from {0.3, 0.5, 0.8}. τ is searched in the set {3, 5, 10, 15, 20}, and τend is searched in {30, 50, 75, 100}. Following (Zhou and Shen 2023), we opt for the early stopping strategy. |