Multi-Target NeuralLss Regression with Learned Confidence Space

Quantifying the uncertainty of model outputs is crucial in improving the reliability of models in multi-target prediction. Therefore, we propose a multi-target joint distribution regression algorithm based on deep neural networks named NeuralLSS, which provides location, scale, and shape information. NeuralLSS utilizes the complete parameters of the multivariate joint distribution as regression targets, and gives the confidence spaces for the targets based on personalized covariance matrices of each sample, thereby quantifying the model’s uncertainty. Experimentation across various real-world multi-target regression datasets reveals that NeuralLSS outperforms all existing distribution regression algorithms in both accuracy and reliability. Compared with previous distribution regression algorithms, we creatively leverage the personalized covariance matrix of each sample to infer efficient and effective confidence spaces, further advancing the performance of distribution regression. This method, therefore, stands as a game-changer in domains that demand reliable multi-target predictions.