Fast Multivariate Spatio-temporal Analysis via Low Rank Tensor Learning

Authors: Mohammad Taha Bahadori, Qi (Rose) Yu, Yan Liu

NeurIPS 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments on both synthetic datasets and real application datasets to demonstrate that our method is not only significantly faster than existing methods but also achieves lower estimation error.
Researcher Affiliation Academia Mohammad Taha Bahadori Dept. of Electrical Engineering Univ. of Southern California Los Angeles, CA 90089 mohammab@usc.edu Qi (Rose) Yu Dept. of Computer Science Univ. of Southern California Los Angeles, CA 90089 qiyu@usc.edu Yan Liu Dept. of Computer Science Univ. of Southern California Los Angeles, CA 90089 yanliu.cs@usc.edu
Pseudocode Yes Algorithm 1 Greedy Low-rank Tensor Learning
Open Source Code No The paper does not include an unambiguous statement that the authors are releasing the source code for their own method, nor does it provide a direct link to a repository for their code.
Open Datasets Yes USHCN: http://www.ncdc.noaa.gov/oa/climate/research/ushcn; CCDS: http://www-bcf.usc.edu/~liu32/data/NA-1990-2002-Monthly.csv; Yelp: http://www.yelp.com/dataset_challenge; Foursquare: [17] X. Long, L. Jin, and J. Joshi. Exploring trajectory-driven local geographic topics in foursquare. In Ubi Comp, 2012.
Dataset Splits Yes For each training length setting, we repeat the experiments for 10 times and select the model parameters via 5-fold cross validation. during the training phase, we use 5-fold cross-validation.
Hardware Specification Yes We measure the run time on a machine with a 6-core 12-thread Intel Xenon 2.67GHz processor and 12GB memory.
Software Dependencies Yes For MTL-L1 , MTL-L21 [19] and MTL-LDirty, we use MALSAR Version 1.1 [27]. We use the MATLAB Kriging Toolbox6 for the classical cokriging algorithms... (footnote 6: http://globec.whoi.edu/software/kriging/V3/english.html)
Experiment Setup Yes For each training length setting, we repeat the experiments for 10 times and select the model parameters via 5-fold cross validation. We split the data along the temporal dimension into 90% training set and 10% testing set. We choose VAR(3) model and during the training phase, we use 5-fold cross-validation. For each dataset, we first normalize it by removing the trend and diving by the standard deviation. Then we randomly pick 10% of locations (or users for Foursquare) and eliminate the measurements of all variables over the whole time span.