Prediction Model Viewer
You use the Prediction Model Viewer to view your Prediction models. The Prediction Model Viewer provides two views for analyzing your Prediction models: Dependency Network view and Decision Tree view.
Use the Prediction Model Viewer to analyze your product dependency network, and then make any changes necessary based on customer behavior on your site.
Dependency Network View
Decision Tree View
Dependency Network View
You use Dependency Network view to analyze the statistical dependencies of the properties in your analysis model. Dependency Network view is the view that first appears when you open a Prediction model in the Prediction Model Viewer. Dependency Network view is a graph of your Prediction model that is represented with nodes and arcs. Nodes correspond to properties and arcs correspond to statistical dependencies among the properties.
.gif "Dependency Network view")
A node appears as an oval with text. A ToolTip appears with the full name of the property if you pause on a node. When no nodes are selected, all nodes appear green in color. Not all nodes are displayed, only the most frequently occurring properties in the data are displayed. For instructions on displaying additional nodes, see Finding All Available Nodes in Dependency Network View.
When you open a Prediction model in Dependency Network view, the nodes are automatically displayed in a manner that is easy to visualize. The nodes continuously move as Commerce Server improves the layout. Using the Dependency Network view you can:
- Stop and restart the improvement process.
- Click a node to highlight the direct relationships between that node and the other nodes in the graph.
- Drag nodes to manually adjust the layout.
For information about working with the layout arrangement of your model, see Arranging the Layout in Dependency Network View. Node colors indicate the following:
- Green. The node is selected.
- Red. The selected node helps to predict this node.
- Blue. This node helps to predict the selected node.
- Purple.****A predictive relationship exists in both directions.
Using the slider along the left side of Dependency Network view you can control the number of relationships exposed based on the strength of the relationship. As the slider is moved down, arcs are removed — that is, the arcs corresponding to the weakest dependencies are removed until only the arcs for the strongest dependences are shown.
You use the slider to help you view a Prediction model with dozens of characteristics and hundreds of relationships because it helps you to focus on the dependencies you are most interested in.
Decision Tree View
You can use the Predictor resource Decision Tree view to analyze the tree components of your Prediction model. You access Decision Tree view by double-clicking a node in Dependency Network view. A decision tree is graphical shorthand for a set of probabilistic rules of the form:
IF X = Value and Y = Value and so on, THEN Target = Value with probability p
The conditions in the "IF" clause of a particular rule correspond to a left-to-right traversal of the tree. In particular, each internal node is annotated with the name of an property, and the out-going edges are annotated with value(s) for that property.
Nodes are oval in shape and purple in color. Leaf nodes, which show the prediction probabilities, are gray rectangular boxes.
.gif "Decision Tree view")
Each leaf node stores a probability distribution for the target property. (A probability distribution is a set of probabilities, one for each possible value of the target.)
- If the target property is categorical and has less than eight values, the probability for each value is depicted in the leaf node as a histogram.
- When a categorical property is binary (that is, the property simply indicates the presence or absence of a value), then only the probability of the property being present is shown as the fraction of the bar that is green (see the Decision Tree view figure).
- For a categorical property with more than eight values, each leaf node is annotated with the most likely value for the target.
- If the target property is continuous, each leaf node is annotated with the expected range of the target.
You can also double-click an internal node to more easily read the path to that node.
You can continue navigating through the tree without closing the dialog box. The dialog box displays the path for the currently selected node.
Note that there is a useful correspondence between the contents of the decision trees and that of the dependency network. Namely, there will be an arc from property A to property B in the dependency network if, and only if, property A is a node in the decision tree for property B. (Both views are saying the same thing: knowing property A helps to predict property B.)
See Also
Copyright © 2005 Microsoft Corporation.
All rights reserved.