MAGMM: A high-dimensional outlier detection algorithm based on a memory-augmented autoencoder and the Gaussian mixture model

Outlier detection is a key research problem in data mining, in which the aim is to identify data objects that deviate significantly from the distribution of the other data. To address the issues of good generalisation by deep autoencoders and the loss of critical information in the original space du...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Information sciences Ročník 721; s. 122574
Hlavní autoři: Zhang, Zhongping, Wang, Kuo, Wang, Zhongman, Li, Junji
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier Inc 01.12.2025
Témata:
Autoencoder
Data mining
Gaussian mixture model
Memory network
Outlier
Outlier
Memory network
Data mining
Autoencoder
Gaussian mixture model
ISSN:0020-0255
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Outlier detection is a key research problem in data mining, in which the aim is to identify data objects that deviate significantly from the distribution of the other data. To address the issues of good generalisation by deep autoencoders and the loss of critical information in the original space during the encoding process, this paper proposes a high-dimensional outlier detection algorithm based on a memory-augmented autoencoder and the Gaussian mixture model. This algorithm enhances the low-dimensional representations generated by the encoder stage of the deep autoencoder through a memory-augmented network, and computes the reconstruction error after memory augmentation, thus amplifying the difference between normal and abnormal data. To preserve the original spatial information of the input data, the algorithm combines the memory-enhanced low-dimensional representations with the reconstruction error features as input for the subsequent estimation network. Moreover, this algorithm jointly learns the deep autoencoder model and Gaussian mixture model, thus significantly reducing the overall loss based on the reconstruction error and sample energy, and consequently detecting outliers in high-dimensional data. Experimental results on several benchmark datasets demonstrate that the proposed algorithm outperforms state-of-the-art outlier detection algorithms, confirming its effectiveness and stability. [Display omitted] •A memory-augmented net amplifying reconstruction error gap between inlier & outliers.•A memory-augmented AE refining latent representations & reconstruction error.•A GMM-based estimation network detecting outliers.
ISSN:0020-0255
DOI:10.1016/j.ins.2025.122574