Transfer Learning for Anomaly Detection through Localized and Unsupervised Instance Selection

Authors: Vercruyssen Vincent, Meert Wannes, Davis Jesse6054-6061

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, LOCIT outperforms a multitude of existing transfer learning and anomaly detection methods on a new transfer learning benchmark for anomaly detection. Moreover, it outperforms its competitors on a real-world anomaly detection problem of identifying anomalous water usage in multiple retail stores.
Researcher Affiliation Academia Vincent Vercruyssen, Wannes Meert, Jesse Davis DTAI group, KU Leuven, Belgium firstname.lastname@kuleuven.be
Pseudocode Yes Algorithm 1: LOCIT(DS, DT , ψ, k)
Open Source Code Yes Finally, we provide an implementation of LOCIT.1 https://github.com/Vincent-Vercruyssen/LocIT
Open Datasets Yes The code, benchmark data, elaborated explanations, full parameter settings, and further experiments are available in an online appendix.2 https://github.com/Vincent-Vercruyssen/LocIT. We start from one of 12 publicly available, multi-class master datasets. To generate a target domain, we sample normal (anomalous) target instances from the largest (second largest) class in the master dataset.
Dataset Splits No The paper mentions "threefold cross-validation" for tuning the SVM's hyperparameters within the LOCIT algorithm, but it does not describe a separate validation set split (e.g., 80/10/10) for the overall model evaluation or hyperparameter tuning of the main anomaly detection task itself.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper mentions using an SVM and KNN, which are algorithms, but does not specify any software dependencies with version numbers (e.g., specific Python libraries or frameworks like scikit-learn or PyTorch versions).
Experiment Setup Yes LOCIT has two hyperparameters. We set the neighborhood size ψ = 20 and k = 10 in SSKNNO. LOCIT tunes the SVMs hyperparameters using threefold cross-validation on the generated training data. It selects either a linear or Gaussian kernel and sets C [0.01, 0.1, 0.5, 1, 10, 100] for both kernels and σ [0.01, 0.1, 0.5, 1, 10, 100] for the Gaussian kernel. We simply use the baselines with the hyperparameters recommended in the original papers or in comparative studies.