198
S. Hu et al.
Table 1. Characteristics of common classifiers
Classifier
Complexity Sensitivity Classification
accuracy
Special
requirements
Linear
regression
Low
Medium
Low
Linear data
Decision tree
Low
High
High
Neglected
correlation
KNN
Medium
High
High
Complex
computation
SVM
Medium
High
Medium
Complex
computation
Naive Bayes
Low
Low
Low
Priori
probability
Artificial
neural
network
High
Low
High
Massive data
calculation of similarity between feature vectors inaccurately affect classification.
The weighted KNN weights the contributions of the P th neighbors and assigns
the larger weights to the nearest neighbors. The problem of imbalanced KNN
sample is effectively solved by this method.
5 Experimental Result
Table 2 contains the average accuracy from four terminals by using one feature
with SNR = 9.07 dB and three classifiers. It is obvious that these features are all
distinguishable with a good identification rate, while constellation side length,
constellation, angle and phase noise spectrum are the best features among them.
Table 2. Classification rate of single feature for QPSK with SNR = 9.07 dB and three
classifiers
QPSK
Bagged tree (%) Weighted KNN (%) Fine Gaussian
SVM (%)
Box dimension
91.9
92.1
77.3
Fractal dimension
89.2
90.2
72.0
Constellation side
length
94.1
94.2
90.1
Constellation diagonal
length
84.1
84.1
52.5
Constellation angle
93.4
93.8
89.7
Phase noise spectrum
93.8
94.1
89.7
S. Hu et al.
Table 1. Characteristics of common classifiers
Classifier
Complexity Sensitivity Classification
accuracy
Special
requirements
Linear
regression
Low
Medium
Low
Linear data
Decision tree
Low
High
High
Neglected
correlation
KNN
Medium
High
High
Complex
computation
SVM
Medium
High
Medium
Complex
computation
Naive Bayes
Low
Low
Low
Priori
probability
Artificial
neural
network
High
Low
High
Massive data
calculation of similarity between feature vectors inaccurately affect classification.
The weighted KNN weights the contributions of the P th neighbors and assigns
the larger weights to the nearest neighbors. The problem of imbalanced KNN
sample is effectively solved by this method.
5 Experimental Result
Table 2 contains the average accuracy from four terminals by using one feature
with SNR = 9.07 dB and three classifiers. It is obvious that these features are all
distinguishable with a good identification rate, while constellation side length,
constellation, angle and phase noise spectrum are the best features among them.
Table 2. Classification rate of single feature for QPSK with SNR = 9.07 dB and three
classifiers
QPSK
Bagged tree (%) Weighted KNN (%) Fine Gaussian
SVM (%)
Box dimension
91.9
92.1
77.3
Fractal dimension
89.2
90.2
72.0
Constellation side
length
94.1
94.2
90.1
Constellation diagonal
length
84.1
84.1
52.5
Constellation angle
93.4
93.8
89.7
Phase noise spectrum
93.8
94.1
89.7