The marketing department of the Adventure Works Cycles company wants to identify the characteristics of previous customers that might indicate whether those customers are likely to buy a product in the future. The AdventureWorks2012 database stores demographic information that describes previous customers. By using the Microsoft Decision Trees algorithm to analyze this information, the marketing department can build a model that predicts whether a particular customer will purchase products, based on the states of known columns about that customer, such as demographics or past buying patterns.
How the Algorithm Works
The Decision Trees algorithm builds a data mining model by creating a series of splits in the tree. These splits are represented as nodes. The algorithm adds a node to the model every time that an input column is found to be significantly correlated with the predictable column. The way that the algorithm determines a split is different depending on whether it is predicting a continuous column or a discrete column.
The Decision Trees algorithm uses feature selection to guide the selection of the most useful attributes. Feature selection is used by all SQL Server Data Mining algorithms to improve performance and the quality of analysis. Feature selection is important to prevent unimportant attributes from using processor time. If you use too many input or predictable attributes when you design a data mining model, the model can take a very long time to process, or even run out of memory. Methods used to determine whether to split the tree include industry-standard metrics for entropy and Bayesian networks*.* For more information about the methods used to select meaningful attributes and then score and rank the attributes, see . A common problem in data mining models is that the model becomes too sensitive to small differences in the training data, in which case it said to be overfitted or over-trained. An overfitted model cannot be generalized to other data sets. To avoid overfitting on any particular set of data, the Microsoft Decision Trees algorithm uses techniques for controlling the growth of the tree. For a more in-depth explanation of how the Microsoft Decision Trees algorithm works, see .
The way that the Microsoft Decision Trees algorithm builds a tree for a discrete predictable column can be demonstrated by using a histogram. The following diagram shows a histogram that plots a predictable column, Bike Buyers, against an input column, Age. The histogram shows that the age of a person helps distinguish whether that person will purchase a bicycle.
The correlation that is shown in the diagram would cause the Microsoft Decision Trees algorithm to create a new node in the model.
As the algorithm adds new nodes to a model, a tree structure is formed. The top node of the tree describes the breakdown of the predictable column for the overall population of customers. As the model continues to grow, the algorithm considers all columns.
Predicting Continuous Columns
When the Microsoft Decision Trees algorithm builds a tree based on a continuous predictable column, each node contains a regression formula. A split occurs at a point of non-linearity in the regression formula. For example, consider the following diagram.
In a standard regression model, you would attempt to derive a single formula that represents the trend and relationships for the data as a whole. However, a single formula might do a poor job of capturing the discontinuity in complex data. Instead, the Microsoft Decision Trees algorithm looks for segments of the tree that are largely linear and creates separate formulas for these segments. By breaking up the data into different segments, the model can do a much better job of approximating the data.
The following diagram represents the tree diagram for the model in the scatterplot above. To predict the outcome, the model provides two different formulas: one for the left branch, with the formula y = .5x x 5, and one for the right branch, with the formula y = .25x + 8.75. The point where the two lines come together in the scatterplot is the point of non-linearity, and is the point where a node in a decision tree model would split.
This is a simple model with only two linear equations; therefore, the split in the tree is immediately after the All node. However, a split can occur at any level of the tree. That means that in a tree containing multiple levels and nodes, where each node is characterized by a different collection of attributes, a formula might be shared across multiple nodes, or apply only to a single node. For example, you might get one formula for a node defined as "customers over a certain age and income", and another in a node that represents "customers who commute long distances". To see the formula for an individual node or segment, just click the node.
Data Required for Decision Tree Models
When you prepare data for use in a decision trees model, you should understand the requirements for the particular algorithm, including how much data is needed, and how the data is used.
The requirements for a decision tree model are as follows:
A single key column Each model must contain one numeric or text column that uniquely identifies each record. Compound keys are not permitted. A predictable column Requires at least one predictable column. You can include multiple predictable attributes in a model, and the predictable attributes can be of different types, either numeric or discrete. However, increasing the number of predictable attributes can increase processing time. Input columns Requires input columns, which can be discrete or continuous. Increasing the number of input attributes affects processing time. For more detailed information about the content types and data types supported for decision tree models, see the Requirements section of .