94
H. K. Palo
Table 2
(continued)
ML algorithms
Attributes
Pros.
Cons.
Support vector machine (SVM) •
Linear discriminant learner
•
Use either RM (Risk Minimization)
or ERM (Empirical RM) to enhance
accuracy in a sample set
•
Accuracy can vary with different
kernel functions
• Can handle large feature set
• Guaranteed convergence with a
unique solution
•
Space complexity
•
Not accurate for a very large feature
set
•
Kernel dependent
•
Unsuitable for variable-length data
Gaussian mixture model
(GMM)
•
Probabilistic approach
•
Uses Expectation Maximization to
speed up response
•
A stochastic classifier that uses
statistical parameters
• Robust
• Computationally efficient
• Faster
• Suitable for a large feature set
• Ease of implementation
• Suitable for global features
•
Text-dependent
•
Can’t avoid exponential functions
•
Not suitable for low dimensional
feature sets
Hidden Markov model (HMM) •
Text-dependent
• Large HMM structures possible using
individual HMMs
•
complex
•
Requires initialization of the suitable
model parameter before training
•
Computationally slow
(continued)