Abstract
In statistical pattern recognition, hidden Markov model (HMM) is the most important technique for modelling patterns that include temporal information such as speech and handwriting. If the temporal information is not taken into account, Gaussian mixture model (GMM) is used. This GMM technique uses a mixture of Gaussian densities to model the distribution of feature vectors extracted from training data. When little training data are available, vector quantisation (VQ) technique is also effective (Tran & Wagner 2002). In fuzzy set theory-based pattern recognition, fuzzy clustering techniques such as fuzzy cmeans and fuzzy entropy are used to design re-estimation techniques for fuzzy HMM, fuzzy GMM, and fuzzy VQ (Tran & Wagner 2000). The first stage in pattern recognition is data feature selection. A number of features that best characterises the considering pattern is extracted and the selection of features is dependent on the pattern to be recognised and has direct impact on the recognition results. The abovementioned pattern recognition methods cannot select features automatically because they treat all features equally. We propose that the contribution of a feature to pattern recognition should be measured by a weight that is assigned to the feature in the modelling process. This method is called soft subspace pattern recognition. There have been some algorithms proposed to calculate weights for soft subspace clustering (Huang et al. 2005, Jing et al. 2007). However a generic framework for the above-mentioned modelling methods has not been built. A generic framework for soft subspace pattern recognition will be proposed in this chapter. A generic objective function will be designed for HMM and maximizing this function will provide an algorithm for calculating weights. Other weight calculation algorithms for GMM and VQ will also be determined from the algorithm for HMM. The proposed soft subspace pattern recognition methods will be evaluated in network intrusion detection. Some preliminary experiments have been done and experimental results showed that the proposed algorithms could improve the recognition rates.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have