| Reconstruction model |
No need for labeled anomaly samples, can capture local and global features of data |
Higher false-positive rate for high-dimensional and large-scale data, requires a large amount of training data to learn data distribution and patterns |
Yoo, Kim & Kim (2019), Zeng et al. (2023a), Xu et al. (2023)
|
| Generative model |
Can model complex, high-dimensional data distributions, can learn data distribution to generate new samples, no need for labeled data |
Training and inference processes are complex and time-consuming for complex data distributions and high-dimensional data, prone to mode collapse which can result in a lack of diversity in generated samples |
Xing, Demertzis & Yang (2020), Talapula et al. (2023), Li et al. (2019)
|
| Predictive model |
Can capture dynamic changes and trends in data, excellent detection performance for time-series data |
Issues with gradient vanishing and exploding, need to continuously adapt to new data distributions for non-stationary data |
Wang et al. (2023), Liu et al. (2021)
|
| Representation learning model |
Suitable for high-dimensional, complex, and large-scale data, better understanding of the intrinsic structure and features of data, can automatically extract useful features |
Requires a large amount of training data and computational resources to train deep neural networks, may have poor interpretability in some cases, prone to overfitting |
Munir et al. (2018), Garg et al. (2019)
|
