Proceedings Article | 17 March 2008
Proc. SPIE. 6974, Multisensor, Multisource Information Fusion: Architectures, Algorithms, and Applications 2008
KEYWORDS: Defense and security, Data mining, Wavelet transforms, Principal component analysis, Detection and tracking algorithms, Visualization, Computing systems, Feature selection, Intelligence systems, Decision support systems
This paper proposes a novel adaptive learning method for data mining in support of decision-making systems. Due to the
inherent characteristics of information ambiguity/uncertainty, high dimensionality and noisy in many homeland security
and defense applications, such as surveillances, monitoring, net-centric battlefield, and others, it is critical to develop
autonomous learning methods to efficiently learn useful information from raw data to help the decision making process.
The proposed method is based on a dynamic learning principle in the feature spaces. Generally speaking, conventional
approaches of learning from high dimensional data sets include various feature extraction (principal component analysis,
wavelet transform, and others) and feature selection (embedded approach, wrapper approach, filter approach, and others)
methods. However, very limited understandings of adaptive learning from different feature spaces have been achieved.
We propose an integrative approach that takes advantages of feature selection and hypothesis ensemble techniques to
achieve our goal. Based on the training data distributions, a feature score function is used to provide a measurement of
the importance of different features for learning purpose. Then multiple hypotheses are iteratively developed in different
feature spaces according to their learning capabilities. Unlike the pre-set iteration steps in many of the existing ensemble
learning approaches, such as adaptive boosting (AdaBoost) method, the iterative learning process will automatically stop
when the intelligent system can not provide a better understanding than a random guess in that particular subset of
feature spaces. Finally, a voting algorithm is used to combine all the decisions from different hypotheses to provide the
final prediction results. Simulation analyses of the proposed method on classification of different US military aircraft
databases show the effectiveness of this method.
