5.9 Bibliography


5.9 Bibliography

Caruana, R. and Hodor, P. (August 20–23, 2000) "High Precision Information Extraction," Sixth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.

Jensen, L, and Martinez, T. (August 20-23, 2000) "Improving Text Classification by Using Conceptual and Contextual Features," Sixth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.

Manning, C. and Schutze, H. (1999) Foundations of Statistical Natural Language Processing, Cambridge, MA: MIT Press.



Chapter 6: Neural Networks: Classifying Patterns

6.1 What Do Neural Networks Do?

Neural networks are software systems modeled after the human process of learning and remembering. They mimic the cognitive neurological functions of the human brain. As such they are capable of predicting new observations from historical samples after executing a process of learning. A neural network can be used to detect a fraudulent transaction, a computer intrusion, and an assortment of other criminal activities, so long as examples of observations are available for training it.

Neural network software comprises programmable memories designed to make predictions. Neural networks were introduced to the marketplace in the mid-1980s and became practical commercial software products only after advances in computing power at the desktop and server level became a reality. They have been used in private industry to do one or more of the following:

  • Classification: discriminating between two things based on similarities, such as separating loan applications into good or bad risks or distinguishing a legal from a fraudulent transaction. They can also be used to discriminate between criminal and legal activities. Chapter 8 illustrates how neural networks are used to detect Internet fraud on an e-commerce site.

  • Clustering: organizing observations into groups with similar features or attributes—for instance identifying groups of customers who buy the same type of products, also referred to as affinity market-basket analysis; they can also be used to group criminals who perpetrate the same types of crimes. This involves using a special type of Kohonen neural network (named after its creator, Dr. Tevo Kohonen). They are also known as self-organizing maps (SOM). Several case studies demonstrating this type of analysis are provided further in this chapter, and book.

  • Generalizing: generalizing from examples about new cases or problems just as humans can learn to model relationships from examples. This same process can be used to model criminal signatures. A case study is provided in this chapter illustrating how neural networks are used to recognize the signature of kerosene in arson investigations.

  • Forecasting: looking at current information and predicting what is likely to happen. Prediction is a form of classification into the future. Neural networks can also be trained on observations in order to predict (with some probability), for example, who is likely to be a smuggler. A demonstration will be provided in this chapter illustrating this process.



6.2 What Is a Neural Network?

Neural network softwares are programs designed to do what human brains have evolved to perform: learn from examples and recognize patterns. As a neural network is trained, it can gradually learn, for example, the patterns of behavior and the attributes of criminals; this is actually done through an adjustment of mathematical formulas that are continuously changing, eventually converging into a model.

This model, in most instances, is the end product of neural networks and is usually in the form of a formula or a series of equations, representing a set of input values and a predictive output. For example, Table 6.1 shows a model designed to predict fraud at an e-commerce Web site. When given the number of visits a person makes to a site (input 1), coupled with a dollar range (input 2) and a particular type of product bought (input 3), a network would then create a formula to compute a probability score (output) for classifying that transaction as being either legal or fraudulent.

Table 6.1: Model with Three Inputs, a Set of Formula Weights, and One Output

Input Types

Values

Network (Weights) Formula

Number of visits

2

Input => *.08766 x

Dollar range

(567–879)

Input => *.00934 x

Product category

4

Input => *.4956 x

  

Output = 68% fraud probability score

Neural networks are software systems designed to do what a human brain does, which is to figure out how to find patterns in the environment in order to survive. Neural networks deal with the basic problems that our ancestors had in hunting, finding paths, knowing the seasons, recognizing prey and predators, and identifying threats to the tribe and village: pattern recognition.