Deep learning requires a large data set. When working with smaller data sets we need an automated approach for feature learning as a pre-processing step to create the training set for a shallow neural network with one hidden layer. The outcome of the neural net classifier is filtered using post-processing rules over a defined wait and quiet period to create an ensemble prediction with higher quality. This gives us the benefits of deep learning, but with a smaller data set.
A partial implementation of this methodology was used for a pilot release for prediction of Common Previsioning Group (CPG) growth failures in 3PAR storage devices, using a trained neural net model with additional derived attributes, pre-processing rules and post-processing steps, applied to 90 days history of logical disk free space for each CPG on the storage device.
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.
Cochran, Dermot; Roe, Martin; Goudar, CV; Martyn, Breda; Das, Mainak; Uppuluri, Pratush; Lennon, Sean; Panda, Basudev; King, Sarah; Dolan, Liam; and Hughes, James Marshall, "A methodology for automated feature engineering in training of shallow neural networks for prediction of multi-linear growth failures (“Nova Algorithm”)", Technical Disclosure Commons, (February 26, 2018)