CourboSpark: Decision Tree for Time-series on Spark

With the deployment of smart meters across many countries, data are being collected at a large scale and volume. These data are collected for billing purposes but also to get analytical insights. Our main goal here is to build an understandable model able to explain the electric consumption patterns regarding several features. We chose to use decision tree models as they are easily comprehensible and have already been parallelized with success. In our industrial context, we often have to work on electrical time-series where the target to predict is neither a label (classification) nor a numerical value (regression) but a time-series representing a load curve. Therefore we use a different split criterion to separate time-series: the inertia. We also need a dedicated method for categorical features since the standard implementation would not work for time-series. This method is based on a hierarchical clustering in order to have a good trade-off between the computational complexity an...

• Type: Journal
• Year: 2015
• Where: PKDD
• Authors: Christophe Salperwyck, Simon Maby, Jérôme Cubillé, Matthieu Lagacherie
• Keywords: Data Mining, PKDD 2015