Sequential deep embedded clustering with dynamic autoencoder

Hasna AlSaeed

doi:10.1007/s11042-026-21953-9

Sequential deep embedded clustering with dynamic autoencoder

2026-11-13

0

PRE

AI

DOI：10.1007/s11042-026-21953-9

Original

Save

Abstract

En 中文

Deep Embedded Clustering (DEC) methods based on Autoencoders have demonstrated strong performance compared to traditional clustering algorithms by jointly learning latent representations and cluster assignments. However, when applied to complex or high-dimensional data, DEC models are prone to Feature-Randomness (FR) and Feature-Drift (FD), which destabilize the learned representations and degrade clustering quality. The Dynamic Autoencoder (DynAE) mitigates this trade-off through a dynamic objective function that gradually shifts the learning focus from reconstruction to clustering. Despite its effectiveness on static data, DynAE is not designed to capture temporal dependencies and therefore struggles with sequential data such as time series or multivariate signals. To address this limitation, we propose a sequential deep embedded clustering framework, called Long Short-Term Memory Dynamic Autoencoder (LSTM-DynAE), which integrates an LSTM-based Autoencoder into the DynAE paradigm. The model is organized into two phases: a pretraining phase, where an LSTM Autoencoder learns temporally aware latent representations, and a clustering phase, where a dynamic loss function progressively transitions from reconstruction-driven exploration to clustering-driven exploitation. This design enables LSTM-DynAE to model sequential data while applying the DynAE-derived dynamic optimization mechanism to manage the FR–FD trade-off. We evaluate LSTM-DynAE on five benchmark sequential datasets and compare it against established sequential and non-sequential deep clustering baselines. Experimental results show that LSTM-DynAE achieves higher clustering performance than the evaluated baselines, reaching up to 97% Accuracy (Acc) and 95% Normalized Mutual Information (NMI), while exhibiting favorable FR-FD behavior during training. These findings indicate that LSTM-DynAE is an effective approach for deep embedded clustering of the evaluated sequential datasets.

Keywords:

Deep embedded clustering

Autoencoder

LSTM

Sequential clustering

Dynamic-autoencoder

Sequential deep embedded clustering with dynamic autoencoder

Abstract

Journal

Organization

Cited Papers