GAT-GAN : A Graph-Attention-based Time-Series Generative Adversarial Network

Generative Adversarial Networks (GANs) have proven to be a powerful tool for generating realistic synthetic data. However, traditional GANs often struggle to capture complex relationships between features which results in generation of unrealistic multivariate time-series data. In this paper, we propose a Graph-Attention-based Generative Adversarial Network (GAT-GAN) that explicitly includes two graph-attention layers, one that learns temporal dependencies while the other captures spatial relationships. Unlike RNN-based GANs that struggle with modeling long sequences of data points, GAT-GAN generates long time-series data of high fidelity using an adversarially trained autoencoder architecture. Our empirical evaluations, using a variety of real-time-series datasets, show that our framework consistently outperforms state-of-the-art benchmarks based on \emph{Frechet Transformer distance} and \emph{Predictive score}, that characterizes (\emph{Fidelity, Diversity}) and \emph{predictive performance} respectively. Moreover, we introduce a Frechet Inception distance-like (FID) metric for time-series data called Frechet Transformer distance (FTD) score (lower is better), to evaluate the quality and variety of generated data. We also found that low FTD scores correspond to the best-performing downstream predictive experiments. Hence, FTD scores can be used as a standardized metric to evaluate synthetic time-series data.

LassoLayer: Nonlinear Feature Selection by Switching One-to-one Links

Along with the desire to address more complex problems, feature selection methods have gained in importance. Feature selection methods can be classified into wrapper method, filter method, and embedded method. Being a powerful embedded feature selection method, Lasso has attracted the attention of many researchers. However, as a linear approach, the applicability of Lasso has been limited. In this work, we propose LassoLayer that is one-to-one connected and trained by L1 optimization, which work to drop out unnecessary units for prediction. For nonlinear feature selections, we build LassoMLP: the network equipped with LassoLayer as its first layer. Because we can insert LassoLayer in any network structure, it can harness the strength of neural network suitable for tasks where feature selection is needed. We evaluate LassoMLP in feature selection with regression and classification tasks. LassoMLP receives features including considerable numbers of noisy factors that is harmful for overfitting. In the experiments using MNIST dataset, we confirm that LassoMLP outperforms the state-of-the-art method.