.png){kind=icon}
.png){kind=icon}
Return
NEURAL NETWORK CUSTOM LOSS FUNCTION FOR FLEXIBLE AND HIGH-QUALITY PREDICTION INTERVALS
.png){kind=icon}
PRE
AI
DOI：10.3934/mfc.2025039.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
. In deep learning research, prediction intervals are increasingly important for a wide range of applications. However, the complexity of today's challenges often demands more flexible and reliable intervals than what traditional methods can provide. To address this, we introduce a custom loss function tailored for deep learning models. Unlike standard approaches, our design directly balances the width of prediction intervals with their coverage, aiming for intervals that are both practical and accurate. The loss function incorporates adjustable penalty coefficients, allowing for precise control over the interval width. When the coefficients are larger, the bounds are pulled closer to the true observations. Smaller values loosen this effect, leading to wider intervals that may favor point accuracy. This trade-off gives the model room to adjust depending on what the application calls for. We evaluated the approach in ten diverse public datasets. The results show that the proposed method consistently produces high-quality prediction intervals and is generalized well. In practice, it demonstrates the ability to adjust smoothly to a variety of scenarios, confirming its potential as a reliable alternative to existing strategies.
Keywords:
Prediction interval
deep learning
custom loss function
penalty coefficients
Journal
M
IF:
0.8
Papers:
21
Citations:
0