6.6 Cost-Sensitive Decision Making
Costs are user-specified numbers that bias Classification. The algorithm uses positive numbers to penalize more expensive outcomes over less expensive outcomes. Higher numbers indicate higher costs.
The algorithm uses negative numbers to favor more beneficial outcomes over less beneficial outcomes. Lower negative numbers indicate higher benefits.
All classification algorithms can use costs for scoring. You can specify the costs in a cost matrix table, or you can specify the costs inline when scoring. If you specify costs inline and the model also has an associated cost matrix, only the inline costs are used. The PREDICTION
, PREDICTION_SET
, and PREDICTION_COST
functions support costs.
Only the Decision Tree algorithm can use costs to bias the model build. If you want to create a Decision Tree model with costs, create a cost matrix table and provide its name in the CLAS_COST_TABLE_NAME
setting for the model. If you specify costs when building the model, the cost matrix used to create the model is used when scoring. If you want to use a different cost matrix table for scoring, first remove the existing cost matrix table then add the new one.
A sample cost matrix table is shown in the following table. The cost matrix specifies costs for a binary target. The matrix indicates that the algorithm must treat a misclassified 0 as twice as costly as a misclassified 1.
Table 6-1 Sample Cost Matrix
ACTUAL_TARGET_VALUE | PREDICTED_TARGET_VALUE | COST |
---|---|---|
0 |
0 |
0 |
0 |
1 |
2 |
1 |
0 |
1 |
1 |
1 |
0 |
Example 6-14 Sample Queries With Costs
The table nbmodel_costs
contains the cost matrix described in Table 6-1.
SELECT * from nbmodel_costs; ACTUAL_TARGET_VALUE PREDICTED_TARGET_VALUE COST ------------------- ---------------------- ---------- 0 0 0 0 1 2 1 0 1 1 1 0
The following statement associates the cost matrix with a Naive Bayes model called nbmodel.
BEGIN dbms_data_mining.add_cost_matrix('nbmodel', 'nbmodel_costs'); END; /
The following query takes the cost matrix into account when scoring mining_data_apply_v
. The output is restricted to those rows where a prediction of 1 is less costly then a prediction of 0.
SELECT cust_gender, COUNT(*) AS cnt, ROUND(AVG(age)) AS avg_age FROM mining_data_apply_v WHERE PREDICTION (nbmodel COST MODEL USING cust_marital_status, education, household_size) = 1 GROUP BY cust_gender ORDER BY cust_gender; C CNT AVG_AGE - ---------- ---------- F 25 38 M 208 43
You can specify costs inline when you invoke the scoring function. If you specify costs inline and the model also has an associated cost matrix, only the inline costs are used. The same query is shown below with different costs specified inline. Instead of the "2" shown in the cost matrix table (Table 6-1), "10" is specified in the inline costs.
SELECT cust_gender, COUNT(*) AS cnt, ROUND(AVG(age)) AS avg_age FROM mining_data_apply_v WHERE PREDICTION (nbmodel COST (0,1) values ((0, 10), (1, 0)) USING cust_marital_status, education, household_size) = 1 GROUP BY cust_gender ORDER BY cust_gender; C CNT AVG_AGE - ---------- ---------- F 74 39 M 581 43
The same query based on probability instead of costs is shown below.
SELECT cust_gender, COUNT(*) AS cnt, ROUND(AVG(age)) AS avg_age FROM mining_data_apply_v WHERE PREDICTION (nbmodel USING cust_marital_status, education, household_size) = 1 GROUP BY cust_gender ORDER BY cust_gender; C CNT AVG_AGE - ---------- ---------- F 73 39 M 577 44