New Dimension Reduction Methods Based on Auto-Encoder for Efficient Clustering

Abstract

Abstract : The rapid growth of time-series data in fields like energy, finance, healthcare, and transportation presents significant analytical challenges due to high dimensionality, temporal dependencies, and limited labeled data. Traditional clustering methods often struggle with these complexities, leading to ineffective identification of underlying groupings. To address these issues, deep clustering methods, particularly those utilizing autoencoders, have become popular. They effectively reduce dimensionality and extract features by compressing high-dimensional data into compact latent representations, capturing essential patterns while minimizing noise. This results in more robust and scalable clustering for complex time-series datasets. However, deep clustering approaches encounter challenges related to Feature Randomness (FR) and Feature Drift (FD), which can destabilize learned features and shift representations during training, negatively impacting clustering quality. Although recent models have attempted to balance the trade-off between FR and FD, they have not adequately addressed the specific characteristics of time-series datasets. To effectively balance the trade-off between FR and FD, this thesis proposes two innovative sequential deep embedded clustering models for time series clustering: the Long Short-Term Memory Dynamic Autoen-coder (LSTM-DynAE) and the Convolutional Long Short-Term Memory Dynamic Autoencoder (ConvLSTM-DynAE). The LSTM-DynAE model integrates LSTM autoencoders with a dynamic loss function that progressively shifts training focus from reconstruction to clustering, maintaining stable and discriminative representations. The ConvLSTM-DynAE extends this framework by incorporating convolution LSTM (ConvLSTM) autoencoders, enabling simultaneous learning of spatial and temporal features, which is particularly beneficial for load profiling in smart grids.