8.14 Assembling the Mining Tools

Using hundreds of these data attributes, several data analyses will be performed, including the construction of several predictive models using an assortment of data mining techniques and tools. We will first perform a visual inspection of fraud transactions and then move on to the construction of predictive models, decision trees, and IF/THEN rules, along with the associated code and final fraud-detection ensemble design. These data mining tools include some of the machine-learning software suites discussed in earlier chapters. They come from such companies as ANGOSS, SAS, SPSS, and others, all of which can be found at the data mining portal, Knowledge Discovery Nuggets (http://www.kdnuggets.com).

These data mining tools incorporate the technologies covered in preceding sections of the book, such as link analysis, SOM, neural networks, and machine-learning algorithms, technologies that are robust and proven. They have been around for years and are very intuitive to use. The objective, however, is not only to understand what the fraud profile looks like, but also to construct predictive models in the form of rules or formulas and code for deterring the criminal activity that takes place at e-commerce sites.

As we have learned, there are two basic types of data mining analyses. One is descriptive and designed to provide some insight into the user, such as a chart, graph, or decision tree, and we will incorporate these types of tools by starting the analysis using a link analysis program and a SOM neural network to discover, view, and explore the hidden associations and patterns related to fraudulent transactions. The second type of data mining analysis consists of creating predictive models. This is where a set of formula weights from a neural network in the form of code or conditional rules is extracted from a sample data set by a machine-learning tool and used to detect and deter fraud. For this second type of modeling analysis, we will use some neural networks and machine-learning algorithm tools. In the end, we will gain an insight into the profile of fraud at a given site and generate code that can be used to detect this criminal activity and for deterring it in real time.