Decision Tree Software

Decision tree classifier is a predictive model that, as its name suggests, is presented in a tree. The following figure shows an example of decision tree. For space efficiency, the tree is drawn left-to-right, as compared to common top-to-bottom representation. The tree is a model that predicts "GENDER" of of employees based on values of "JOBPOST", "SALARY", "RACE", etc. In another words, given values of "JOBPOST", "SALARY", "RACE", etc. of an employee, the tree predicts "GENDER" of the employee.

Decision tree software.

The left most node "WHOLE" is the root of the tree. At the second level, the root node is divided into five sub-nodes based on the values of the variable "JOBPOST". Note that areas of rectangles indicate gender value distribution in nodes. Green is for "Male" and red is for "Female". The root node shows rather even gender distribution, i.e., 51.5% is male and the rest female. After the split at the second level, nodes show more biases towards a certain gender. The middle three nodes are heavily populated with males. The first node does not show much change in gender distribution. So it is further divided based on "RACE". Further dividing the node "RACE=Ethnic" does not yield improvement in gender distribution. So the node no longer divided. On the other hand, dividing the node "RACE=White" based on the salary boundary at $30,000 results in sub-nodes rich with female and male populations respectively. Similarly, the bottom node of the second level is divided twice, producing terminal nodes rich in female and male populations respectively.

Terminal nodes represent predictions or decisions. Decisions at terminal nodes are made based on most frequent categorical values, in this case, "GENDER". If a terminal node has more male population, decision is made to "Male". Otherwise to "Female". Decision trees having terminal nodes dominated by a single class lead to higher accuracy. A perfect decision tree with prediction 100% accuracy will consist of terminal nodes having population of the same categorical values. For example, either all males or all females. Normally this ideal distribution does not exist. So accuracy of decision trees falls!

Tree Splitting Criteria - The Cramer effect!

Decision trees are constructed by dividing tree nodes repeatedly. Splitting of a node involves selection of a dividing variable whose values are used in designing branches. Splitting variables are selected automatically according to a user-selected quality measurement criterion. The following is a list of splitting criteria that CMSR Cramer Decision Tree supports;

  • Cramer index (Exclusively supported only on CMSR).
  • Entropy and entropy gain ratio.
  • GINI diversity index.
  • Chi-square statistic and probability.
  • Expected accuracy or expected probability.
  • Twoing.
  • Manual selection (for arbitrary segmentation analysis).

The quality of predictive models is measured in terms with accuracy on un-seen data. Note that un-seen data means data not used in developing the decision tree. There are two factors that we can measure: accuracy and numbers of nodes. The latter is a very important factor for un-seen data. For the same or similar accuracy, smaller numbers of nodes mean that trees were constructed with more general terms (or splitting values). Thus they can work better with un-seen data. Note that higher numbers imply that trees use splitting variables with large numbers of values. This can result in negative impact on un-seen data.

Cramer decision tree is an exclusive feature!

In general, Cramer decision trees tend to have smaller numbers of nodes. It is noted that most node splitting criteria inherently favor splits with many branches. This tends to result in trees with a large number of nodes. Cramer leverages this phenomena with "degrees of freedom". As a consequence, it tends to produce most compact decision trees. It is noted that less branching nodes means more large sized nodes, which translates to higher statistical supports. Generally, this will lead to better predictive accuracy on unseen data. Therefore, Cramer trees are considered to be generally bests and exclusively supported in CTM. For comparisons, read Cramer Tree.

Decision tree as a Segmentation tool

Decision tree divides populations into smaller segments repeatedly. At a node, it selects a single variable in such a way that values of the variable boost proportions of the largest categorical value in each resulting segments. If the population is insurance policies, each segmentation will try to increase the proportion of either "never-claimed" or "claimed" customers. This tends to lead segments with higher portion of claimed policies. Similar analogy applies to other application areas such as credit, finance, direct mail catalog responses, customer churns, and so on. There are many applications that this type of hierarchical segmentation is useful.

Probablity tree indicating probability 
distribution of drill-down nodes.

Statistical information of nodes is shown at the right-hand side of the window. It includes pie charts and histograms. For the insurance example, reds may represent "claimed" customer portions and greens for "never-claimed" customers. Nodes in red indicate that over 50% customers of the segments have claims. Green nodes have less than 50% of claim customers. In addition to red nodes, nodes with lower height green bar may be of interest. Note that they represent relatively higher proportions of risky customers. As you can see from the figure, it's very easy to identify segments using various statistical gauges of StarProbe/CramerTree. Visualization includes the following information;

  • Most frequent categorical values.
  • Probabilities of a selected categorical value.
  • Averages and totals of numerical fields.
  • Distribution: pies and histograms.
Marketing Response Analysis with Gains & Profit Charts

Breakdown of populations into smaller segments induces segments having concentration of certain values. This provides means for cost-effective customer selection methods. For example, for catalog marketing, segments can be visualized with a special response and profit gains chart as shown at the right figure. The blue curve indicates response capture ratios. The green curve shows quantity (or volume) capture ratios. Similarly, the red curve describes profit and loss amounts. Steep rising curves at the left end are a good indication of good segmentation, since selecting a small number of left-end customer segments can capture most potential responses. This will lead to a small number of catalog mailings, and therefore result in efficient marketing campaigns. For more, read Direct Mail Catalog Marketing.

decision tree classifier

Decision Tree Segmentation for RFM Marketing?

Decision tree is very useful as a segmentation tool for Direct Mail Marketing and customer segmentation. Decision tree segmentation is especially well suited for RFM Marketing for customer targeting.

For information about software, please read Data Mining Software. Software download is available from the page.