Oblique Random Forest via Regularized Multisurface Proximal Support Vector Machine
Reshma Rastogi, Anubhav David · 2019 Global Conference for Advancement in Technology (GCAT) · 2019
We propose a novel approach to train oblique random forests using a linear classifier where the feature axis is not invariably orthogonal to the decision hyperplanes at each internal node of the base model. At each non-leaf node, for the multiclass classification problem, we group all the training samples into two groups of classes corresponding to the underlying geometric characteristics with respect to a randomly chosen feature subspace. Subsequently, two discriminatory hyperplanes are obtained, as per each class, by employing the Regularized Multisurface Proximal Support Vector Machine (RegMPSVM). While optimizing the impurity criteria, the test hyperplane of the corresponding internal node is selected from either of the bisectors of the two hyperplanes of the two groups of classes. As the tree propagates, to tackle the small sample size issue, various regularisation approaches are employed. Using this method, the oblique random forests are grown. Proposed framework's potency is exhibited by comparison with the state-of-the-art frameworks on 40 real-world benchmark classification datasets of miscellaneous research fields. In the expression of prediction accuracy, the ascendancy of our proposed framework is demonstrated by the classification results.