Predictor Schema Example
This example shows all the related tables and entries for a model configuration named PurchaseCfg1.
PredictorModelCfgs
| ModelCfgName | SiteName |
|---|---|
| PurchaseCfg1 | Retail |
This example assumes that the PurchaseCfg1 model configuration consists of three sources of data:
Demographic information from the User table (additional rows not shown).
| UserID | Age | Gender | Education |
|---|---|---|---|
| jilluser@roguecellars.com | 23 | F | 19 |
| barneyuser@arborshoes.com | 49 | M | 3 |
| jackuser@frogkick.com | 28 | M | 13 |
Product Purchase information from the Purchases table (additional rows not shown).
| UserID | SKU | QTY |
|---|---|---|
| jilluser@roguecellars.com | wine_fine_au_zin_19 | 3 |
| jilluser@roguecellars.com | tv_big_36_m3 | 1 |
| jilluser@roguecellars.com | pumps_bl_6 | 1 |
Ad Click information from the AdClicks table (additional rows not shown). If Click is set to 1, the ad was seen but not clicked. If Click is set to 2, the ad was seen and clicked.
| UserID | Ad | Click |
|---|---|---|
| jilluser@roguecellars.com | www.thewinecellar.com/ad_zin12 | 1 |
| jilluser@roguecellars.com | www.thewinecellar.com/ad_cab3 | 2 |
| jilluser@roguecellars.com | www.arborshoes.com/ad4c | 1 |
PredictorDataTables
| ModelCfgName | TableName | Type | Case Column |
Pivot Column |
Aggregate Column |
Aggregate Type |
|---|---|---|---|---|---|---|
| PurchaseCfg1 | User | 0 | UserID | <NULL> | <NULL> | <NULL> |
| PurchaseCfg1 | Purchases | 1 | UserID | SKU | QTY | 0 |
| PurchaseCfg1 | AdClicks | 1 | UserID | Ad | Click | 1 |
| PurchaseCfg1 | PurchaseCfg1_Attributes | 2 | "N/A" | <NULL> | <NULL> | <NULL> |
When data from the three tables (User, Purchases, and AdClicks) are combined for analysis, a left join is done with User as the master table, on the UserID column. That is, each user from the User table defines a case. If there are users represented in either the Purchases or AdClicks table that are not represented in the User table, they will not be included in the analysis. In general, if there is a dense table (there can be, at most, one), then it is the master table in the left join. If all tables are sparse, then the first table listed is the master table for the left join. Case column names do not have to be identical, but their values must be correlated.
The first row in the following attributes table specifies all of the SKUs to have a Discrete and Modeled As Binary distribution and specifies that SKUs should be used to predict other properties and not to be predicted. The second row specifies to use the Age property to predict other properties, and not to be predicted. This means that no Decision Tree that predicts the Age or SKU properties will be built. The row also specifies the Distribution to be Continuous, Lognormal and not Modeled As Binary. The third row specifies all of the ads are to have a distribution of type Advertisement and that ads are to be predicted but not used for prediction.
The following table is an example of a model configuration that will predict ads based on SKU and Age.
PredictorAttributes_PurchaseCfg1
| Prop ID |
Parent ID |
Name | Table Name |
Column Name |
Distri- bution |
UseTo Predict |
Predict |
|---|---|---|---|---|---|---|---|
| 1 | -1 | SKU | <NULL> | <NULL> | 4 | True | False |
| 2 | -1 | Age | <NULL> | <NULL> | 2 | True | False |
| 3 | -1 | AD | <NULL> | <NULL> | 8 | False | True |
The following table shows a model built from the PurchaseCfg1 model configuration.
PredictorModels
| Model Name |
ModelCfg Name |
Model Type |
Date Created |
Build Time |
Measured Accuracy Sample Fraction |
Measured Accuracy Max Predictions |
K |
|---|---|---|---|---|---|---|---|
| Purchase1 | PurchaseCfg1 | 0 | getdate() | 30000 | .05 | 10 | <NULL> |
(PredictorModels continued)
| Max Buffer Size |
Input Attribute Fraction |
Output Attribute Fraction |
Sample Size |
Complexity Penality |
Minimum CasesTo Split |
Data Fit Score |
Recommend Score |
Data |
|---|---|---|---|---|---|---|---|---|
| <NULL> | 1 | 1 | -1 | 100 | 5 | 11.08 | 0.3946 | <Binary> |
Copyright © 2005 Microsoft Corporation.
All rights reserved.