To construct a high-quality classifier one has to have high-quality data. No method of classifier construction, whether based on neural networks or a statistical one, will result in a classifier good enough if the given set of examples is not full and representative for the problem that the system would have to work with.
Solving the problem of classification is one the most important implementations of neural networks.
The problem of classification is a task of ranging a pattern in one of several pairwise disjoint sets. Examples of these tasks can be determination of a bank client's creditworthiness, medical tasks requiring estimation of a disease outcome, tasks of investment management (selling/buying shares or holding them back depending on the market situation), task of working out viable companies and those susceptible to bankruptcy.
Purpose of classifier
Solving problems of classification requires ranging the available static patterns (parameters of the market situation, medical examination data or information about a client) in certain classes. There can be several methods of data representation. The prevalent one is a method when a pattern is represented by a vector. Components of the vector represent various parameters of the pattern that influence the decision of assigning the pattern to a class .
To construct an Algorithm of a classifier based on neural network , it is required to follow the steps .
1. Work with the data. Form a database with samples typical for the given task.
2. Decompose the whole set of data into two sets: training and testing sets (decomposition in three sets is also possible: training, testing and confirmation sets).
To make the Preliminary processing of a high-quality classifier -
1. Select a system of attributes typical for the task and appropriately transform the data for feeding to the network input (normalization, standardization etc.) It is desirable to obtain linearly separable space of the patterns set.
2. Select a system for coding the output values (classical coding, 2 by 2 coding etc.)
To Construct, train and evaluate the network quality ,following are the steps --
1. Select the network topology: number of layers, number of neurons in layers etc.
2. Select a function of neuron activation (e.g., «sigmoid')
3. Select an algorithm of training the system.
4. Evaluate the quality of the network functioning, basing on confirmation sets or other criteria. Optimize the architecture (decrease weights, reducing dimensionality of the attributes space).
5. Settle on the variant of the network that provides the best ability to generalize. Evaluate the quality using the testing set.
Usage and diagnostics
1. Find out the levels of various factors that influence the decision (heuristic approach).
2. Make sure the network provides necessary accuracy of classification (number of incorrectly recognized samples is low).
3. If necessary return to step 2 after changing the method of samples representation or database.
4. Proceed to practical usage of the network for solving the problem.
Here i have given some research tips for the implementation and processing requirements of a classifier based on neural network .
About Author / Additional Info: