LB-CGM: Latent Based Conditional Generative Model with Reliable Distribution Prediction

Abstract

Randomness exists either due to the inherent noise of the problem or lack of important input features, which could lead to multimodality of the data distribution. Therefore, in more and more scenarios, it is required not only to predict a single point-value, but also the distribution of the prediction. However, well-studied prediction models usually focus on point prediction that minimizes the mean squared error or the mean absolute error. These approaches could miss important knowledge when their outputs are applied to the downstream decision process. In this paper, we combine the advantages of both GANs (Generative Adversarial Nets) and VAEs (Variational Auto-Encoders), and introduce a latent-based conditional generative model (LB-CGM) to handle the distribution regression problems. The VAE framework is adopted, and the adversarial network is applied to estimate the validity of the generated sample. Besides, the latent-based reconstruction loss is introduced to mitigate mode collapse, in which the direct pairwise comparison between the original and generated samples ensures the correctness and completeness of the generated mode pattern. In this work, we explore a path for the generative model to be used in probabilistic prediction problems. This method can produce conditional prediction distribution close to the actual distribution and is verified on both the synthetic dataset and benchmark dataset.