Beyond Graph Convolution: Multimodal Recommendation with Topology-aware MLPs
Authors: Junjie Huang, Jiarui Qin, Yong Yu, Weinan Zhang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on three real-world datasets verify TMLP s superiority over nine baselines. We also find that by discarding the internal message passing in GCNs, which is sensitive to node connections, TMLP achieves significant improvements in both training efficiency and robustness against existing models. |
| Researcher Affiliation | Academia | Junjie Huang, Jiarui Qin, Yong Yu, Weinan Zhang* Shanghai Jiao Tong Univeristy, Shanghai, China EMAIL |
| Pseudocode | No | The paper describes the methodology in prose and with a framework diagram, but it does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions: "We implement TMLP with Py Torch using MMRec (Zhou 2023), a comprehensive repository for multimodal recommendation models." This indicates the use of a third-party repository (MMRec) for implementation, but does not state that the code for TMLP itself is open-sourced or provided. |
| Open Datasets | Yes | We conduct experiments on three categories from the Amazon review dataset (He and Mc Auley 2016a; Mc Auley et al. 2015): Baby, Sports, and Electronics. |
| Dataset Splits | Yes | To ensure fair comparison, we use the same evaluation settings and 8:1:1 data split as in (Zhou et al. 2023) on the filtered 5-core data. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. It only discusses training epochs and efficiency without specifying the underlying hardware. |
| Software Dependencies | No | The paper mentions using "PyTorch" and "MMRec (Zhou 2023)", but it does not specify version numbers for PyTorch or any other library/solver used, which is required for reproducibility. |
| Experiment Setup | Yes | We set the embedding size to 64, initialized with Xavier (Glorot and Bengio 2010). Optimal hyperparameters are found via grid search, with learning rates in {1e-4, 5e-4, 1e-3, 5e-3}, MLP layers in {2, 3, 4}, NA loss weight α in Equation (13) from 0 to 2 (interval 0.1), and sampling size K from 3 to 10. The regularization weight λ is set to 0. We fix the hidden dimension of MLP as 512, activation function as tanh and dropout rate as 0. The ratio of visual features βm in Equation (2) is set to 0.1, with r = 1 in Equation (8). Training is capped at 1000 epochs with early stopping at 20 consecutive epochs. |