6th European Conference - Academic Conferences

Kesav Kancherla and Srinivas Mukkamala
coefficients as 12. Our method performed best when we used 50-50 compression setting. In 50-50
there is more noise added due to compression. As our method is based on using this noise for
detection we obtained better accuracy in 50-50 setting. In next section we explain the model selection
process and Receiver Operation Characteristic (ROC) curves.
5.1 Model selection for SVMs
In any predictive learning task, such as classification, both model and parameter estimation method
should be selected in order to achieve a high level of performance of the learning machine. Recent
approaches allow a wide class of models of varying complexity to be chosen. Then the task of
learning amounts to selecting the sought-after model of optimal complexity and estimating parameters
from training data (Cherkassy, 2002: 109-133; Lee and Lin 2000). Within the SVMs approach, usually
parameters to be chosen are (i) The penalty term C which determines the trade-off between the
complexity of the decision function and the number of training examples misclassified; (ii) The
mapping function and (iii) The kernel function such that . In the case of RBF kernel, the width,
which implicitly defines the high dimensional feature space, is the other parameter to be selected
(Chang and Lin, 2001). Figures 2 and 3 gives the model graph obtained during training.
Figure 2: Model graph obtained during training of SVM obtained for YASS at block size 9,
compression rates 50-50 and coefficients used for embedding equal to 12
Figure 3: Gives model graph obtained during training SVM for YASS at block size 14, compression
rates 75-75 and coefficients used for embedding equal to 10
148