LGMRec: Local and Global Graph Learning for Multimodal Recommendation

Authors: Zhiqiang Guo, Jianjun Li, Guohui Li, Chaoyang Wang, Si Shi, Bin Ruan

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments conducted on three benchmark datasets demonstrate the superiority of our LGMRec over various state-of-the-art recommendation baselines, showcasing its effectiveness in modeling both local and global user interests.
Researcher Affiliation Academia 1 School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China 2 School of Software Engineering, Huazhong University of Science and Technology, Wuhan, China 3 Wuhan Digital Engineering Institute, Wuhan, China 4 Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen, China {zhiqiangguo, jianjunli, guohuili}@hust.edu.cn, sunwardtree@outlook.com, shisi@gml.ac.cn, binruan0227@gmail.com
Pseudocode No The paper describes the methodology using mathematical equations and textual descriptions but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes We implement LGMRec2 with MMRec (Zhou 2023). 2https://github.com/georgeguo-cn/LGMRec
Open Datasets Yes To evaluate our proposed model, we conduct comprehensive experiments on three widely used Amazon datasets (Mc Auley et al. 2015): Baby, Sports and Outdoors, Clothing Shoes and Jewelry. We refer to them as Baby, Sports, Clothing for brief.
Dataset Splits Yes For each dataset, we randomly split historical interactions into training, validation, and testing sets with 8 : 1 : 1 ratio.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments, such as GPU/CPU models or memory.
Software Dependencies No The paper mentions implementing LGMRec with MMRec (Zhou 2023), but does not specify other software dependencies like Python version, PyTorch/TensorFlow versions, or other libraries with their version numbers.
Experiment Setup Yes For a fair comparison, we optimize all models with the default batch size 2048, learning rate 0.001, and embedding size d = 64. For all graph-based methods, the number L of collaborative graph prorogation layers is set to 2. In addition, we initialize the model parameters with the Xavier method (Glorot and Bengio 2010). For our model, the optimal hyper-parameters are determined via grid search on the validation set. Specifically, the number of modal graph embedding layers and hypergraph embedding layers (K and H) are tuned in {1, 2, 3, 4}. The number A of hyperedge is searched in {1, 2, 4, 8, 16, 32, 64, 128, 256}. The dropout ratio ρ and the adjust factor α are tuned in {0.1, 0.2, . . . , 1.0}. We search both the adjust weight λ2 of contrastive loss and the regularization coefficient λ1 in {1e 6, 1e 5, . . . , 0.1}. The early stop mechanism is adopted, i.e., the training will stop when R@20 on the verification set does not increase for 20 successive epochs.