Anomaly Attribution with Likelihood Compensation

Authors: Tsuyoshi Idé, Amit Dhurandhar, Jiří Navrátil, Moninder Singh, Naoki Abe4131-4138

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We apply our approach to a real-world building energy prediction task and confirm its utility based on expert feedback.
Researcher Affiliation Industry Tsuyoshi Id e, Amit Dhurandhar, Jiˇr ı Navr atil, Moninder Singh, Naoki Abe IBM Research, T. J. Watson Research Center {tide, adhuran, jiri, moninder, nabe}@us.ibm.com
Pseudocode Yes Algorithm 1 summarizes the iterative procedure for finding δ.
Open Source Code No The paper does not provide an explicit statement or link to open-source code for the described methodology. It discusses open-source tools as baselines but does not state that their own implementation is open-sourced.
Open Datasets No The paper mentions using "Boston Housing data (Belsley 1980)" and "energy consumption data for an office building in India" but does not provide concrete access information (link, DOI, repository, or explicit statement of public availability with author/year citation for the latter) for these datasets. While Boston Housing is a well-known dataset, the paper does not provide a direct access link or specific citation for it.
Dataset Splits Yes We held out 20% of the data as Dtest (Ntest = 101), and trained a random forest on the rest.
Hardware Specification No The paper mentions running a comparison for SV+ on a "laptop PC (Core i7-8850H)" but does not specify the hardware used for the main experiments with LC.
Software Dependencies No The paper does not specify version numbers for any software dependencies.
Experiment Setup Yes For the ℓ1 parameter, we gave ν = 0.1 for LC, then chose ν = 0.005 for LIME+, so that LIME+ has on average the same number of nonzero elements as LC. The ℓ2 parameter λ was chosen as 0.5 for LC and LIME+ to have approximately the same scale. For the learning rate κ, in our experiments, we fixed κ = 0.1 and shrank it (geometrically) by a factor of 0.98 in every iteration. In the experiment, we fixed N s = 1000 following (Ribeiro, Singh, and Guestrin 2016) and ηi = 1 for all i after standardization.