ODIM: Outlier Detection via Likelihood of Under-Fitted Generative Models
Authors: Dongha Kim, Jaesung Hwang, Jongjin Lee, Kunwoong Kim, Yongdai Kim
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To validate the superiority and efficiency of our method, we provide extensive empirical analyses on close to 60 datasets. |
| Researcher Affiliation | Collaboration | Dongha Kim* 1 Jaesung Hwang* 2 Jongjin Lee 3 Kunwoong Kim 4 Yongdai Kim 4 1Department of Statistics and Data Science Center, Sungshin Women s University, Seoul, Republic of Korea 2SK Telecom, Seoul, Republic of Korea 3Samsung Research, Seoul, Republic of Korea 4Department of Statistics, Seoul National University, Seoul, Republic of Korea. |
| Pseudocode | Yes | We provide the ODIM s pseudo algorithm in Algorithm 1. |
| Open Source Code | Yes | The implementation code for our method is publicly available at https://github.com/jshwang0311/ODIM. |
| Open Datasets | Yes | We analyze 57 benchmark datasets for OD covering tabular, images, and texts, all of which are sourced from ADBench2 (Han et al., 2022a). |
| Dataset Splits | No | The paper discusses 'training data' and mentions using a 'currently estimated generative model' for early stopping based on Wasserstein distance, which serves as a form of validation for model updates. However, it does not specify a distinct 'validation dataset split' from the original data for hyperparameter tuning or general model evaluation, only referring to 'training data' for performance assessment. |
| Hardware Specification | Yes | We utilize the Pytorch framework to run our algorithm using a single NVIDIA TITAN XP GPU. |
| Software Dependencies | No | The paper mentions using 'Pytorch framework' and 'Adam optimizer' but does not specify their version numbers. It also refers to 'BERT' or 'RoBERTa', which are models, not specific software dependencies with version numbers for the implementation. |
| Experiment Setup | Yes | We use two hidden layered DNN architectures for building the encoder and decoder and set K, the number of samples drawn from the encoder used for constructing the IWAE objective function, to 50. We minimize the IWAE objective function in (1) with the Adam optimizer (Kingma & Ba, 2014) with a mini-batch size of 128 and a learning rate of 5e-4. To run the ODIM, we fix the two hyper-parameters, Nu and Npat, to 10. For ensemble learning, we train 10 pairs of encoder and decoder, each of which is trained independently. |