E.R. Jones (USA)
Data Mining, Neural Networks, Java, C, classification, IMSL
In the past, business analytics were mainly confined to applying computationally simple algorithms such as moving averages and decision rules to solve forecasting and classification problems. Today this is changing. More than ever, organizations require better forecasting accuracy and lower classification error rates than can be achieved using simple methods. Many are turning to using neural network algorithms to power their forecasting and decision making applications. These applications range from managing inventory and customer service to text mining and pattern recognition. Unfortunately these more sophisticated algorithms are also more computationally intensive. Training a neural network solution can take minutes or hours, even on a high speed computer. Software engineers must take care to ensure that these applications provide practical as well as correct solutions. Languages such as Java are popular choices for programming business analytic applications, but an important software design question is whether Java and similar languages have the computational power needed to implement computationally intensive business analytics. This paper describes research characterizing the relationship between the complexity of neural network solutions and the computing demands of neural network solutions. Rules of thumb and benchmarks are provided for designing business analytic applications in Java and C using neural networks. These can be used to identify applications requiring parallel solutions and off-line network training.
Important Links:
Go Back
