Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Unifying Reconstruction and Density Estimation via Invertible Contraction Mapping in One-Class Classification
Authors: Xiaolei Wang, Tianhong Dai, Huihui Bai, Yao Zhao, Jimin XIAO
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on tabular data, natural image, and industrial image data demonstrate the effectiveness of our method. Code is available at URD. 4 Experiment |
| Researcher Affiliation | Collaboration | Xiaolei Wang1,2 Tianhong Dai4 Huihui Bai3 Yao Zhao3 Jimin Xiao1 1Xi an Jiaotong-Liverpool University 2University of Liverpool 3Beijing Jiaotong University 4Applied Computing Technologies |
| Pseudocode | No | The paper describes methods using mathematical equations and textual descriptions, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at URD. As mentioned in the Abstract, our code and data will be made publicly available. |
| Open Datasets | Yes | For tabular AD, we select 12 benchmark datasets from OODS [36] and ADBench [20], which include diverse fields, such as healthcare, finance, and social sciences, etc. For visual AD tasks, we conduct experiments on both natural and industrial datasets: CIFAR-10 and MVTec AD [5], respectively. |
| Dataset Splits | Yes | Following previous OCC works [38, 57], only one class in CIFAR-10 is selected as the normal class, while the remaining classes are considered abnormal. Following MCM [53], all datasets are implemented with identical dataset partitioning. The training set is denoted as Itrain = {Ii, yi}N1 i=1, where each Ii is a normal sample with label yi = 0, and N1 is the number of samples. The testing set is denoted as Itest = {It i, yt i}N2 i=1, which contains both normal and abnormal samples, yt i {0, 1}, and N2 is the number of testing samples. |
| Hardware Specification | Yes | All experiments are conducted on a single NVIDIA RTX 3090 24GB GPU. |
| Software Dependencies | No | The implementation is based on Py Torch. |
| Experiment Setup | Yes | The implementation is based on Py Torch. In the tabular AD task, we use a three-layer Multilayer Perceptron (MLP) with Leaky Re LU activation function as a feature extractor. The flow-based structure is modeled using K = 2 and R = 3. In particular, the hyperparameters are set to ̸ = 0.5 and ̸ = 0.01 across 20 tabular datasets. During training, Adam optimizer is used to update the network, where learning rate and weight decay are 1 10 5 and 1 10 7, respectively. For natural image datasets, CIFIAR-10, we use the same convolutional neural network (CNN) as DO2HSC [57] to extract the deep representation of each image. Same as the tabular setting, we set K = 2, R = 3, ̸ = 0.5, and ̸ = 0.001. The learning rate and weight decay are set to 2 10 5 and 1 10 7, respectively. We train the model for 50 epochs with a batch size of 32. For the industrial dataset MVTec, a frozen Wide Res Net-50 [23] is used as the backbone for both the encoder and decoder. For fair comparison, each image is resized to 224 224 without any data augmentation. We select K = 3, R = 3, ̸ = 0.5, and ̸ = 0.001 in the framework. During training, the learning rate and the weight decay are selected as 5 10 4 and 1 10 6, respectively. We train the model for 100 epochs with a batch size of 16. |