13.07.2015 Views

Thesis - Instituto de Telecomunicações

Thesis - Instituto de Telecomunicações

Thesis - Instituto de Telecomunicações

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

98 CHAPTER 5. FEATURE SELECTION AND CLASSIFICATION3. Embed<strong>de</strong>d methods - in this class of feature selection methods the classifier is anintegral part of the feature selection process, therefore the features to be used are<strong>de</strong>ci<strong>de</strong>d during the learning stage [22].We implemented a feature selection based on a wrapper approach, using the SequentialForward Search feature selection procedure in a user tuned scenario.5.2.1 Feature Selection ImplementationFor the purpose of feature selection we consi<strong>de</strong>r that we have a classifier system that receivesa subset of features and returns the EER of the system.Features selection is therefore<strong>de</strong>signed in the context of user authentication. The employed classifiers will be explainedin the next sections.As a first approach, we searched for a set of features that could be used for all users. Weused a greedy search algorithm, called Sequential Forward Search [206, 217], schematically<strong>de</strong>scribed in algorithm 2.Algorithm 2: Sequential Forward Search Feature Selection.Input: feature-vector/* feature-vector the complete feature vector */1 last-eer ← 12 feature-subset-vector ← []3 while true do4 vector-eer ← [0, ··· , 0] nf5 for i =1...n f do6 feature-test-vector ← feature-vector i ∪ feature-subset-vector7 vector-eer i = TestClassifier(feature-test-vector)8 end9 if Min(vector-eer) > last-eer thenReturn(feature-subset-vector)endbest-feature ← ArgMin(vector-eer)feature-subset-vector ← best-feature ∪ feature-subset-vectorlast-eer ← Min (vector-eer)end101112131415We found that creating a different feature subset for each user improved the systemrecognition performance. We therefore applied the same greedy search algorithm to find

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!