Fast Unbalanced Optimal Transport on a Tree

Authors: Ryoma Sato, Makoto Yamada, Hisashi Kashima

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We confirm the effectiveness of the proposed method through numerical experiments. We run the experiments on a Linux server with an Intel Xeon CPU E7-4830 @ 2.00 GHz and 1 TB RAM. We aim to answer the following questions. (Q1) How fast is the proposed method? (Q2) How accurately can tree GKR approximate Euclidean GKR? (Q3) Is the proposed method applicable to large-scale datasets? We also investigate the noise robustness of GKR, but we defer this to the appendices
Researcher Affiliation Academia Ryoma Sato1,2 Makoto Yamada1,2 Hisashi Kashima1,2 1Kyoto University 2RIKEN AIP {r.sato@ml.ist.i, myamada@i, kashima@i}.kyoto-u.ac.jp
Pseudocode Yes Algorithm 1 describes the pseudo code of the algorithm.
Open Source Code Yes Our code including a C++ implementation and Python wrapper for our algorithm is available at https://github.com/joisino/treegkr.
Open Datasets Yes We use Chicago Crime dataset2, where each measure corresponds to a day and contains Dirac mass in the place where a crime occurred in that day. [...] We use New York taxi dataset3 from November 1, 2015 to January 31, 2016.
Dataset Splits Yes The scale is determined by 10 training data so that the relative error is minimized (see Appendix E).
Hardware Specification Yes We run the experiments on a Linux server with an Intel Xeon CPU E7-4830 @ 2.00 GHz and 1 TB RAM. We also conduct the same experiments with a laptop computer with an Intel Core i5-7200U @ 2.50 GHz and 4 GB RAM.
Software Dependencies No The paper mentions using the Lemon graph library and a C++ implementation of the Sinkhorn algorithm, but does not provide specific version numbers for these software components.
Experiment Setup Yes For each n = 27, 28, . . . , 220, we generate 10 random trees with n nodes. The amount of mass in each node is an integer drawn from an i.i.d. uniform distribution from 0 to 106. The weight of each edge is also drawn from the same distribution. [...] We compute the GKR distance GKReuc of these measures in the Euclidean space exactly with λd = λc = λ [10 4, 10.0]. We then compute the GKR distance GKRtree of these measures using the quadtree of depth 15 without random translation.