Pattern recognition methods feature selection

Pattern recognition methods have been used for the description of air pollution in the industrialized region at the estuary of the river Rhine near Rotterdam. A selection of about eight chemical and physical-meteorological features offers a possibility for a description that accounts for out 70% of the information that is ccmprised in these features with two parameters only. Prediction of noxious air situations scmetimes succeeds for a period of at most four hours in advance. Seme-times, hewever, no prediction can be made. Investigations pertaining to the correlation between air conpo-sition and complaints on bad smell by inhabitants of the area show that, apart frem physical and chemical descriptors, other features are also involved that depend on human perception and bdiaviour. 

Second, the number (d) of features must be much greater than that (n) of patterns. Otherwise, a classifier may be found that even separates randomly selected classes of the training set. A ratio of n/d >3 is acceptable, n/d > 10 is desirable (Chapter 1.6, details in Chapter 10.4). This second reason forces the user of pattern recognition methods into feature selection. In almost all chemical applications of pattern recognition is the number of original raw features too large and a reduction of the dimensionality is necessary. 

Additional application of chemical knowledge to the selection of features or to the classifier construction has improved the classification results C1933. A comparison between pattern recognition methods and a sophisticated interpretative library search system for mass spectra ( STIRS C39, 4221) has indicated some superiority of the STIRS-system C172, 202, 3321. A decision tree pattern recognition was recommended by Neisel et. al. C2051 as a supplement to library search. 

Another presumptive application of pattern recognition techniques is the interpretation of quality control data- Two such problems were described by Kowalski C147]- Quality control measurements were performed to characterize production items of an explosive in one case and of beryllium parts in the other. Pattern recognition methods may be useful for the evaluation of specification limits. If no classification method is capable of distinguishing between good and poor items, then it may be supposed that the data (quality control measurements) are insufficient-Feature selection and combination of features may indicate those measurements which provide most information on the separation into good and poor items C1483. 

Chapters 9 and 10 deal with preprocessing of original data and feature selection. These problems must be treated at the beginning of a pattern recognition application. These Chapters have not been positioned at the beginning of the text because a more detailed description of these subjects requires some basic knowledge of pattern recognition methods. 

D. Coomans, M.P. Derde, D.L. Massart and I. Broeckaert, Potential methods in pattern recognition. Part 3 Feature selection with ALLOC, Anal. Chim. Acta, 133, 241-250 (1981). 

Feature selection is the process by which the data or variables liq>or-tant for class assignment are determined. In this step of a pattern recognition study the various methods differ considerably. In the hyperplane methods, the strategy is to begin with a block of variables for the classes, calculate a classification function, and test it for classification of the training set. In this initial phase, generally many more variables are included than are necessary. Variables are then detected in a stepwise process and a new rule is derived and tested. This process is repeated until a set of variables is obtained that will give an acceptable level of classification. 

Unsupervised feature selection methods are also studied in close relation with unsupervised clustering algorithms. In this case, the goal is to find an optimal subset of features with which clusters are well identified. In pattern recognition, researchers want to find a subset of dimensions with which they can better detect specific patterns in a dataset. 

Although feature selection has been one of the most intensively investigated areas of pattern recognition no general theory exists up to now. Therefore/ the chemist should never forget the chemical background of his classification problem. A feature selection based on chemical knowledge may often be much more effective than mathematical methods. 

Interpretation of low resolution mass spectra is the field with the greatest number of applications of pattern recognition techniques in chemistry. Numerous methods of preprocessing, feature selection, training, and evaluation have been tested with mass spectral data in about 100 papers. Probably the first application of pattern recognition ideas in mass spectrometry has been reported by Raznikov and Talroze C235J this Russian paper is summarized in C224J. 

Pudil, P., Novovicova, J. and Kittler, J. (1994). Floating search methods in feature selection. Pattern recognition letters, IS, pp. 1119-1125. 

Feature extraction is a process where the raw SEMG signal is represented into a feature vector which is then used to separate the desired output, e.g. different hand grip postures. The success of the ECS based on pattern recognition depends on the selection and extraction of features [6]. To get a reasonable processing time, the recorded SEMG data were pre-processed. There are various methods to do this but the most widely used is data segmentation because it can improve the accuracy and the response time to the controller. 

Pudil, P., Novovicov, J., Kittier, J. Floating Search Methods in Feature Selection. Pattern Recognition Letters 15,1119-1125 (1994)... 

---