Tree-Sliced Variants of Wasserstein Distances

Authors: Tam Le, Makoto Yamada, Kenji Fukumizu, Marco Cuturi

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we evaluated the proposed TSW kernel k TSW (Equation (5)) for comparing empirical measures in word embedding-based document classification and topological data analysis.
Researcher Affiliation Collaboration Tam Le RIKEN AIP, Japan tam.le@riken.jp Makoto Yamada Kyoto University & RIKEN AIP, Japan makoto.yamada@riken.jp Kenji Fukumizu ISM, Japan & RIKEN AIP, Japan fukumizu@ism.ac.jp Marco Cuturi Google Brain, Paris & CREST ENSAE cuturi@google.com
Pseudocode Yes Algorithm 1 Partition_Tree_Metric(s, X, xs, h, HT)
Open Source Code Yes We have released code for these tools2. 2https://github.com/lttam/Tree Wasserstein.
Open Datasets Yes We evaluated k TSW on four datasets: TWITTER, RECIPE, CLASSIC and AMAZON, following the approach of Word Mover s distances [39], for document classification with SVM.
Dataset Splits Yes For SVM, we randomly split each dataset into 70%/30% for training and test with 100 repeats, choose hyper-parameters through cross validation, choose 1/t from {1, q10, q20, q50} where qs is the s% quantile of a subset of corresponding distances, observed on a training set, use one-vs-one strategy with Libsvm [12] for multi-class classification, and choose SVM regularization from 10 2:1:2 .
Hardware Specification Yes We ran experiments with Intel Xeon CPU E7-8891v3 (2.80GHz), and 256GB RAM.
Software Dependencies No The paper mentions 'Libsvm [12]' but does not provide specific version numbers for software dependencies needed for replication.
Experiment Setup Yes For SVM, we randomly split each dataset into 70%/30% for training and test with 100 repeats, choose hyper-parameters through cross validation, choose 1/t from {1, q10, q20, q50} where qs is the s% quantile of a subset of corresponding distances, observed on a training set, use one-vs-one strategy with Libsvm [12] for multi-class classification, and choose SVM regularization from 10 2:1:2 .