Feature Selection

Ecological data sets are often high dimensional: data sets derived from satellite imagery may contain hundreds of intercorrelated spectral bands. It is desirable to train from a reduced set of attributes, to reduce both the size of the model-fitting search space, and the risk of over-fitting the data.

There are two common approaches: filters and wrappers. Filter methods pre-analyze the data to select a suitable subset of the available features - for example, principal component analysis may be used to generate and select a small number ofhighly influential features for training. However, the model-learning phase may uncover higher-order interactions which are invisible to a filter approach; these are better handled by wrapper approaches, which 'wrap' an optimization algorithm around the underlying model-learning algorithm, repeatedly training from different feature sets until a good feature set is found. Evolutionary algorithms are frequently used as the wrapper.

Solar Power

Solar Power

Start Saving On Your Electricity Bills Using The Power of the Sun And Other Natural Resources!

Get My Free Ebook

Post a comment