Embedded Files
Ordinal Classification
Ordinal Classification
Ordinal classification is a classification of data labels on an ordinal scale of discrete values. Numerical difference between two classes may not have same meaning hence can not be used as nominal classification. Treating ordinal classification as numerical continues requires making invalid assumptions that equal numerical differences in class have equal meaning.
Providing an ideal separation of classes
Providing an ideal separation of classes
- Classes must be properly ordered
- Distance between classes must be non-trivial (not arbitrarily close to 0)
- Require or apply no distance assumptions, restricting relative class distances
- Ideally resulting in homogeneous class clusters
Ordinal Hyperplane Loss (OHPL) with 3 Classes
Ordinal Hyperplane Loss (OHPL) with 3 Classes
OHPL addresses the unique requirements of the ordinal classification by using a point specific large margin loss function to group classes, while directly adhering to the ordinal information that is represented in the data.
The framework provides improved results while avoiding the addition of complexity to redesign the problem." in the "Ordinal Hyperplane Loss (OHPL) with 3 Classes
OHPL: Hyperplane Normal Line
OHPL: Hyperplane Normal Line
- Solid line Hyperplane Centroids (HCs) are the mean hyperplane values for the class; require minimum 1 unit separation
- Movement perpendicular to HCs, ensures class (HC) separation while minimizing point distances lead to improved point classification
OHPL: Movement Parallel to Hyperplanes
OHPL: Movement Parallel to Hyperplanes
- Movement parallel to the HC (blue arrows) does not lead to additional separation of classes or improved point classification
- Data point distances are calculated as the L1 distance from their respective HC aligning it to a MAE
Recurrent Neural Networks (RNN's)
Recurrent Neural Networks (RNN's)
RNN is a class of artificial neural networks where connections between nodes form a directed graph along a temporal sequence.
Cons of RNN:
Cons of RNN:
- RNN suffers from memory loss. If text is too long, it's relevant information might be forgotten during earlier training steps.
- During back-propagation, RNN suffer from vanishing gradient issue.
Both of these issues result in lack of learning.
Long Short Term Memory (LSTM)
Long Short Term Memory (LSTM)
LSTM is similar to RNN but the operations within LSTM's cell distinguishes them from simple RNN
- Cell state and gates within LSTM allows it to keep or forget information. Cell state carry information through the processing of a sequence.
- During this process information gets added or removed to the cell state via various gates. These gates learn to keep or forget information during the training.
OHPL and LSTM
OHPL and LSTM
LSTM (Artificial Recurrent Neural Networks), built using OHPL function, performs on par with existing state-of-the-art ordinal classifiers on small data sets, while demonstrating vastly improved results on larger standard benchmark data sets.
Page updated
Google Sites
Report abuse