Resource-efficient decision tree-based ensemble classifiers with reduced memory footprint, low training time, and low classification latency.
This project offers a framework to augment standard decision-tree based ensemble classifiers resulting in reduced memory footprint, low training time, and low classification latency. The key idea in achieving these properties is to use a two-step strategy: the first step is to train a small model that is sufficient to classify the majority of the queries correctly. The second step involves identifying specific subsets of the training data and train secondary expert models for these fewer harder cases where the small model is at high risk of making a classification mistake.
The project includes two different classifiers:
(1) RADE: Resource-efficient classifier for supervised anomaly detection.
(2) Duet: Resource-efficient multiclass classifier.