|
Regarding the issue of abnormal sales in online shopping websites, the utilization of the MAD-GAN model based on LSTM is considered for handling it, owing to its adeptness in detecting temporal data patterns. The MAD-GAN employs Gan Loss (cross-entropy loss function), which possesses issues of asymmetry and gradient vanishing. To address this, an enhancement using Wasserstein Loss is proposed (referred to as WLoss), resulting in a model termed MAD-WGAN. WLoss asserts that the optimal state is achieved when the data distribution Q aligns perfectly with the sample distribution P. This implies treating the data distribution Q as the true distribution R. However, a certain disparity exists between the actual data distribution Q and the true distribution R. Consequently, a novel definition of optimal distance, termed Wasserstein Deviation Loss, is introduced. Wasserstein Deviation Loss posits that the best state is achieved when a minor difference exists between data distribution Q and sample distribution P (regulated by a parameter β). This state is labeled as WDLoss. To assess the effects of these improvements, a MAD-WDGAN model employing WDLoss is proposed and compared with MAD-GAN and MAD-WGAN. Across the same test dataset, MAD-WDGAN outperforms MAD-GAN and MAD-WGAN models by 5.42% and 0.45%, respectively, in terms of accuracy. This suggests the advantageous nature of WDLoss.
|
