Tensorial Change Analysis Using Probabilistic Tensor Regression

Authors: Tsuyoshi Idé3902-3909

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments As discussed, the problem of tensorial change analysis is new and existing methods are not directly comparable about the change analysis part. Thus, we focus on 1) demonstrating the practical utility of BALS in computing mode-wise change analysis scores for a real-world application. We also illustrate major features of BALS by 2) comparing with alternatives on metrics such as computational time. Synthetic data To show general features of BALS in comparison to the alternatives, we synthetically generated mode-3 (M = 3) tensor data in (d1, d2, d3) = (10, 8, 5) with a given set of the coefficients and randomly generated covariance matrices... As summarized in Fig. 2, in spite of many outliers due to the t-noise, BALS outperformed the alternatives in RMSE.
Researcher Affiliation Industry Tsuyoshi Id e IBM Research, Thomas J. Watson Research Center tide@us.ibm.com
Pseudocode Yes Algorithm 1 Bayesian ALS (BALS) for Tensor Regression.
Open Source Code No The paper does not provide concrete access to source code for the described methodology.
Open Datasets Yes Next, we tested outlier detection capabilities of BALS using publicly available London School data (Goldstein 1991) to pick unusually well-performing schools in the data cleansing task, in which schools whose median of exam.score is greater than 25 are defined as outliers.
Dataset Splits Yes The parameters are optimized using 5-fold cross-validation (CV), so the root mean squared error (RMSE) was minimized.
Hardware Specification No The paper does not provide specific details about the hardware used for running its experiments.
Software Dependencies No We used R s tensr package for d-GPR. This mentions a software package but without a specific version number.
Experiment Setup Yes Following (Kohn, Smith, and Chan 2001), we fix α0 = 1, β0 = 10 6 so the prior becomes near noninformative. In BALS, we used an approximation φl,rφl,r φl,r φl,r in evaluating λ and Σl,r for numerical stability. In BALS, we picked R = 7 that gave the maximum AUC value. We picked R = 4 that maximized AUC.