A tangent few-shot anomaly detection framework based on two-stage embedding representation for tabular data

Few-shot anomaly detection (FSAD) aims to identify anomalies using limited samples. While this task has been widely studied in the image field, existing image-based models cannot be directly applied to tabular data. The reason is that the low-level detailed information in tabular data is crucial for few-shot anomaly detection and cannot be overlooked, as it often is when dealing with images. Existing tabular-based few-shot anomaly detection methods rely on generative models to increase the sample size for training. However, these generative models struggle to extract the correct distribution from small amounts of data and suffer from severe overfitting. To address these issues, this paper proposes a few-shot anomaly detection framework for tabular data, called Maxout Low-dimentional Tangent Data Description (MLTDD). Specifically, we design a Maxout Two-Stage Low-Dimensional Representation Extractor (MaxLR) for low-dimensional representation, which trains a random model in the first stage and embeds the data into a low-dimensional space using a non-probabilistic model. Moreover, we propose a few-shot one-class model called Tangent Support Vector Data Description (TSVDD) for anomaly detection. We conduct experiments on ten datasets and show that MLTDD achieves an overall improvement of more than 10% compared to advanced baseline models and generalizes well in the face of unknown anomalies. The ablation experiment also proved the correctness of our constructed module.