ATTENTION-GUIDED DEEP ADVERSARIAL TEMPORAL SUBSPACE CLUSTERING FOR MULTIVARIATE SPATIOTEMPORAL DATA

Francis Ndikum Nji, Vandana Janeja, Jianwu Wang

Keywords: graph attention transformer, generative models, agenerative models, time distributed convlstm2d, deep unsupervised clustering, u-net autoencoders, multidimensional multivariate spatiotemporal climate data, adversarial models.

Abstract: Deep subspace clustering models provide an efficient solution to the problem of unsupervised subspace clustering of non-linear high-dimensional spatiotemporal data. These clustering solutions are often needed in applications such as motion segmentation, tracking ice sheet dynamics, and detecting anomalous melting events where high-dimensional data exhibit complex temporal dependencies and lie on multiple non-linear manifolds that cannot be effectively captured by traditional linear clustering methods. Existing deep clustering models learn non-linear mappings by projecting data into a latent space in which data lie in linear subspaces and exploit the self-expressiveness property. While this approach has shown impressive performance, they have shortcomings. First, they employ "shallow" autoencoders that completely rely on the self-expressiveness of latent features and disregard potential clustering errors. Second, they focus solely on global features while overlooking local features in subspace self-expressiveness learning. Third, they do not capture long-range dependencies or positional information, both of which are crucial for effective spatial and temporal feature extraction and often lead to sub-optimal clustering outcomes. Fourth, their application to 4D multivariate spatiotemporal data remains underexplored. To address these limitations, we propose a novel Attention-Guided Deep Adversarial Subspace Clustering (A-DATSC) for multivariate spatiotemporal data. A-DATSC incorporates a deep subspace clustering generator and a quality-verifying discriminator that work in tandem. Inspired by the U-Net architecture, the generator preserves the spatial and time-wise structural integrity, reduces the number of trainable parameters, and improves generalization through the use of stacked TimeDistributed ConvLSTM2D layers. The generator introduces a graph attention transformer-based self-expressive network which captures local spatial relationships, global dependencies, and both short- and long-range correlations crucial for understanding how distant regions and time periods influence each other. When evaluated on three real-world multivariate spatiotemporal datasets, A-DATSC outperforms state-of-the-art shallow and deep subspace clustering models with significant margins.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 8040