Using MineSet for Knowledge Discovery

Trying to apply a visualization to data without prepro-cessing can often lead to incomprehensible results because of cluttered and confused displays. A solution is to use data mining to extract just the interesting results and to visualize them. There are two basic ways to carry out the mining: supervised and unsupervised. For the supervised case, the user has a particular goal in mind, while in the second case the user is performing a global search for interesting patterns. MineSet provides two classifiers for goal directed learning. A classifier is a model for predicting one attribute of a set of data given the values for several other attributes. Prior to inducing a classifier a user must select one of the attributes as a label (the goal). The classifier is constructed based on a training set where the label values are known to be accurate. Once built, the model may be applied to new data where the label is unknown. The Evidence Visualizer displays the classifier structure produced by running an evidence inducer on the serv-er. Figure 4 shows the Evidence Visualizer being used to determine data attributes correlate with earning more than $50,000 in salary. Each attribute is a row of pie charts. Each pie chart represents an interval or value of the attribute. The attributes are listed in order of usefulness for predicting the label (salary in this case). The height of the pies show the number records having a particular value. Interaction with the scene allows the user to ask questions like 'what is the probability that someone earns over $50,000 given that their education is High School graduate and their occupation is professional specialty?'. The Tree Visualizer is used to display the decision tree structure produced by a decision tree inducer. The attribute shown at the root of the tree is the one determined to be the most important. After the root split, branches are further split based on other attributes until further splits are deemed to be insignificant statistically. At this point leaf nodes are close to containing only one label value. In figure 5 one can see that odor is the most important attribute for determining whether or not a mushroom is poisonous. It completely determines the edibility, except when there is no odor, in which case stalk shape is used to discriminate further. A common commercial application for this is determining loan approvals based on …