Anomaly Detection for Tabular Data with Internal Contrastive Learning
Authors: Tom Shenkar, Lior Wolf
ICLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments show that our method leads by a sizable accuracy gap in comparison to the literature and that the same default rule of hyperparameters selection provides state-of-the-art results across benchmarks. |
| Researcher Affiliation | Academia | Tom Shenkar & Lior Wolf Blavatnik School of Computer Science, Tel Aviv University EMAIL,EMAIL |
| Pseudocode | No | No pseudocode or algorithm blocks were found. |
| Open Source Code | Yes | The full implementation of our method and scripts for reproducing the experiments are attached as a supplementary zip file. This archive includes a README file and a list of requirements that support seamless reproducibility. |
| Open Datasets | Yes | The second set employs the Multi-dimensional point datasets from the Outlier Detection Data Sets (ODDS)1. It contains 31 datasets, including two of the four datasets above. 1http://odds.cs.stonybrook.edu/, accessed January 2021 |
| Dataset Splits | No | Following Zong et al. (2018); Bergman & Hoshen (2020), the training set contains a random subset of 50% of the normal data. The test set contains the rest of the normal data, as well as all the anomalies. No explicit mention of a validation set split was found. |
| Hardware Specification | Yes | The google colab infrastructure was used to run the experiments. GPU: Tesla K80, 12GB GDDR5 VRAM; CPU: Single core Xeon Processors @2.3Ghz; RAM: 24 GB. For the GOAD baseline that required larger memory, we used a 32GB GPU and 512GB RAM. |
| Software Dependencies | No | The paper mentions a "list of requirements" in a supplementary zip file but does not explicitly list specific software dependencies with version numbers in the main text. |
| Experiment Setup | Yes | We fix τ = 0.01 and u = 200. The value of u, which is often much larger than k, provides enough capacity throughout all experiments, without the need to tune it for each problem. We set k proportionally to the input dimension d. For d smaller than 40, we set k = 2, for d in the range [40, 160] we employ k = 10, and for d > 160, k takes the value d 150. Training employs the Adam optimizer (Kingma & Ba, 2014) with a learning rate of 10 3 . |