Get room level accuracy using DRED dataset

[bontorhumala]

• [x] Get smart meter features (max mean) - modify pcasvmconf feature extractor
• [x] Get room level ground truth - create adapter that will generate format required by classifier
• [x] Filter out timestamp that doesnt have ground truth
• [x] NILMTK on DRED to get appliances power consumption
• [x] Use Orange to create appliance association rule
• [x] Create grouping rule
• [x] Integration + measure performance

[bontorhumala]

Room level ground truth and smart meter features is in dred_adapter

[bontorhumala]

Accuracy is so low due to class imbalance (1 vs 4 because only 1 room is occupied at a time).

Reference in choosing accuracy matrics:

• http://stats.stackexchange.com/questions/12702/what-are-the-measure-for-accuracy-of-multilabel-data
• https://github.com/scikit-learn/scikit-learn/pull/4287
• http://scikit-learn.org/stable/modules/generated/sklearn.metrics.precision_recall_fscore_support.html

Reference in over/undersampling:

• http://contrib.scikit-learn.org/imbalanced-learn/auto_examples/index.html

[bontorhumala]

See dred branch