MetricEmbedding: Accelerate Metric Nearness by Tropical Inner Product

Authors: Muyang Cao, Jiajun Yu, Xin Du, Gang Pan, Wei Wang

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that our method achieves up to 60 speed improvements over state-of-the-art approaches, and efficiently scales from 1e4 1e4 to 1e5 1e5 matrices with significantly lower memory usage. [...] Table 1. Comparison of Methods: Ours, TRF, HLWB, and PAF for Different Matrix Sizes. The table compares computation time, NMSE, and triangle inequality violations for four methods: Ours, TRF, HLWB, and PAF. Figure 2. Performance comparison of the model under different missing rates and noise levels.
Researcher Affiliation Academia 1College of Computer Science and Technology, Zhejiang University 2Shanghai Innovation Institute 3Shool of Software Technology, Zhejiang University 4 The Hong Kong University of Science and Technology (Guangzhou) 5 The Hong Kong University of Science and Technology. Correspondence to: Xin Du <EMAIL>, Wei Wang <EMAIL>.
Pseudocode Yes Algorithm 1 Training Procedure for Metric Embedding Input: Target matrix D, Adam W optimizer,threshold Output: Optimized weight matrices W0, W1, . . . , Wd Optimized bias vectors b1, . . . , bd Final output matrix O [...] Algorithm 2 Minibatch Training Procedure for Metric Embedding Input: Target matrix D, batch size B, Adam W optimizer, threshold Output: Optimized weight matrices W0, W1, . . . , Wd Optimized bias vectors b1, . . . , bd Final output matrix approximation from minibatches
Open Source Code No The recently open-sourced Python C implementation of HLWB may offer improved runtime performance.
Open Datasets Yes We used the Cora dataset, which contains 2708 nodes, and followed the standard data split used in GCN (Kipf & Welling, 2017).
Dataset Splits Yes We used the Cora dataset, which contains 2708 nodes, and followed the standard data split used in GCN (Kipf & Welling, 2017).
Hardware Specification Yes The experiments are conducted on a machine with an RTX 4090D GPU (24GB) and 32 v CPUs (Intel Xeon Platinum 8474C), running on a Linux-based operating system.
Software Dependencies No For Metric Embedding, the learning rate is set to 0.001 using the Adam W optimizer. The model is initialized as described earlier. We use Python implementations of TRF and HLWB, and the Julia implementation of PAF.1 All methods are evaluated based on their convergence time.
Experiment Setup Yes For Metric Embedding, the learning rate is set to 0.001 using the Adam W optimizer. The model is initialized as described earlier. [...] We utilized the dropout adj function in Py G to randomly perturb edges with a perturbation ratio of 0.1 to generate augmented views. The configuration used a learning rate of 0.01, 1000 epochs, with hidden and projection dimensions set to 64.