Data-Informed Geometric Space Selection
Authors: Shuai Zhang, Wenqi Jiang
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate the proposed method on real-world datasets to demonstrate its capability in dealing with practical tasks, including personalized ranking and link prediction for relational graphs. |
| Researcher Affiliation | Academia | Shuai Zhang ETH Zurich cheungshuai@outlook.com Wenqi Jiang ETH Zurich wenqi.jiang@inf.ethz.ch |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described. |
| Open Datasets | Yes | We conduct experiments on two well-known datasets, Movie Lens 100K and Movie Lens 1M. [...] We use two datasets including WN18RR [8, 16] and FB15K-237 [8, 16] for model evaluation. |
| Dataset Splits | Yes | We hold 70% entries in each user s interactions as the training set, 10% entries as the validation set for model tuning, and the remaining 20% for model testing. [...] WN18RR: 86, 835/3, 034/3, 134 training/validation/test triples. FB15K-237: 272, 115/17, 535/20, 466 training/validation/test triples. |
| Hardware Specification | No | The paper does not provide specific hardware details used for running its experiments. |
| Software Dependencies | No | The paper mentions using Adam as an optimizer, but does not specify any software dependencies with version numbers. |
| Experiment Setup | Yes | For all methods, the total dimension PN i=1 bi is set to 100 for a fair comparison to ensure the same model size. The curvatures for spherical and hyperbolic models are set to 1 and -1, respectively. N is set to 5. K is tuned among {1, 2, 3, 4}. Regularization rate is chosen from {0.1, 0.01, 0.001}. m is fixed to 0.5. Adam is adopted as the optimizer. [...] Learning rate is tuned amongst {0.01, 0.005, 0.001}. For all experiments, we report the average over 5 runs. We set the kernel size to 5 and stride to 3 for convolution operation in the gating network. N is set to 5 and K is tuned among {1, 2, 3, 4}. The number of negative samples (uniformly sampled) per factual triple is set to 50. Optimizer Adam is used for model learning. We perform early stopping if the validation MRR stops increasing after 10 epochs. |