Decision Thresholds and Profit Charts

There are two distinct ways of using decision processing in Enterprise Miner:
  • Making firm decisions in the modeling nodes and comparing models on profit and loss summary statistics. For this approach, you include all possible decisions in the decision matrix. This is the traditional approach in statistical decision theory.
  • Using a profit chart to set a decision threshold. For this approach, there is an implicit decision (usually a decision to "do nothing") that is not included in the decision matrix. The decisions made in the modeling nodes are tentative. The profit and loss summary statistics from the modeling nodes are not used. Instead, you look at profit charts (similar to lift or gains charts) in the Model Comparison node to decide on a threshold for the do-nothing decision. Then you use a Transform Variables or SAS Code node that sets the decision variable to "do nothing" when the expected profit or loss is not better than the threshold chosen from the profit chart. This approach is popular for business applications such as direct marketing.
To understand the difference between these two approaches to decision making, you first need to understand the effects of various types of transformations of decisions on the resulting decisions and summary statistics.
Consider the formula for the expected profit of decision d in case i using (without loss of generality) revenue and cost:
A(i,d) = Sum(i)Q(i,t,d)Post(i,t) = Sum(i)[Revenue(t,d) – Cost(i,d)]Post(i,t) = Sum(i)Revenue(t,d)Post(i,t) – Cost(i,d)Sum(t)Post(i,t)
Now transform the decision problem by adding a constant to the tth row of the revenue matrix and a constant ci to the ith row of the cost matrix, yielding a new expected profit A'(i,d):
A’(i,d) = Sum(t)[Revenue(t,d) + r(sub-t)]Post(i,t) – [Cost(i,d) + c(sub-i)]Sum(t)Post(i,t) = A(i,d) + Sum(i)r(sub-i)Post(i,t) + c(sub-i)
In the last expression above, the second and third terms do not depend on the decision. Hence, this transformation of the decision problem will not affect the choice of decision.
Consider the total profit before transformation and without adjustment for priors:
Total Profit = Sum(i)F(i)C(i) = Sum(i)F(i)Q(i,T(i), D(i)) = Sum(i)F(i)[Revenue(T(i), D(i)) – Cost(i, D(i))]
After transformation, the new total profit, TotalProfit', is:
Total Profit’ = Sum(i)F(i)[Revenue(T(i), D(i)) + r(sub-i) – Cost(i,D(i)) – c(sub-i)] = Sum(i)F(i)[r(sub-i) – c(sub-i)]
In the last expression above, the second term does not depend on the posterior probabilities and therefore does not depend on the model. Hence, this transformation of the decision problem adds the same constant to the total profit regardless of the model, and the transformation does not affect the choice of models based on total profit. The same conclusion applies to average profit and to total and average loss, and also applies when the adjustment for prior probabilities is used.
For example, in the German credit benchmark data set (SAMPSIO.DMAGECR), the target variable indicates whether the credit risk of each loan applicant is good or bad, and a decision must be made to accept or reject each application. It is customary to use the loss matrix:
Customary Loss Matrix for the German Credit Data
Target Value
Decision
Accept
Reject
Good
0
1
Bad
5
0
This loss matrix says that accepting a bad credit risk is five times worse than rejecting a good credit risk. But this matrix also says that you cannot make any money no matter what you do, so the results might be difficult to interpret (or perhaps you should just get out of business). In fact, if you accept a good credit risk, you will make money, that is, you will have a negative loss. And if you reject an application (good or bad), there will be no profit or loss aside from the cost of processing the application, which will be ignored. Hence, it would be more realistic to subtract one from the first row of the matrix to give a more realistic loss matrix:
Realistic Loss Matrix for the German Credit Data
Target Value
Decision
Accept
Reject
Good
– 1
0
Bad
5
0
This loss matrix will yield the same decisions and the same model selections as the first matrix, but the summary statistics for the second matrix will be easier to interpret.
Sometimes a decision threshold K is used to modify the decision-making process, so that no decision is made unless the maximum expected profit exceeds K. However, making no decision is really a decision to make no decision or to "do nothing." Thus. the use of a threshold implicitly creates a new decision numbered Nd+1. Let Dk(i) be the decision based on threshold K. Thus:
D(sub-k)(i) = arg max(d=1, N(sub-d)) A(i,d) if A(i,d) > K, = N(sub-d) + 1 otherwise.
If the decision and cost matrices are correctly specified, then using a threshold is suboptimal, since D(i) is the optimal decision, not Dk(i). But a threshold-based decision can be reformulated as an optimal decision using modified decision and cost matrices in several ways.
A threshold-based decision is optimal if "doing nothing" actually yields an additional revenue K. For example, K might be the interest earned on money saved by doing nothing. Using the profit matrix formulation, you can define an augmented profit matrix Profit* with Nd+1 columns, where:
Profit*(t,d) = Profit(t,d) where d <= N(sub-d), = K where d = N(sub-d) + 1
Let D*(i) be the decision based on Profit*, where:
D*(i) = arg max(d=1, N(sub-d)) Sum(t)Profit*(t,d)Post(i,t)
Then D*(i) = DK(i). Equivalently, you can define augmented revenue and cost matrices, Revenue*> and Cost*, each with Nd+1 columns, where:
Revenue*(t,d) = Revenue(t,d) where d <= N(sub-d), = K where d = N(sub-d) + 1
Cost*(i,d) = Cost(i,d) where d <= N(sub-d), = –K where d = N(sub-d) + 1
Then the decision D*(i) based on Revenue* and Cost* is:
D*(i) = arg max(d=1, N(sub-d)) Sum(t) Profit*(t,d)Post(i,t)
Again, D*(i) = DK(i).
A threshold-based decision is also optimal if doing anything other than nothing actually incurs an additional cost K. In this situation, you can define an augmented profit matrix Profit* with Nd+1 columns, where:
Profit*(t,d) = Profit(t,d) – K where d <= N(sub-d), = 0 where d = N(sub-d) + 1
This version of Profit* produces the same decisions as the previous version, but the total profit is reduced by K Sum F(i) regardless of the model used. Similarly, you can define Revenue* and Cost* as:
Revenue*(t,d) = Revenue(t,d) where d <= N(sub-d), = K where d = N(sub-d) + 1
Cost*(i,d) = Cost(i,d) – K where d <= N(sub-d), = 0 where d = N(sub-d) + 1
Again, this version of the Revenue* and Cost* matrices produces the same decisions as the previous version, but the total profit is reduced by K Sum F(i) regardless of the model used.
If you want to apply a known decision threshold in any of the modeling nodes in Enterprise Miner, use an augmented decision matrix as described above. If you want to explore the consequences of using different threshold values to make suboptimal decisions, you can use profit charts in the Model Comparison node with a non-augmented decision matrix. In a profit chart, the horizontal axis shows percentile points of the expected profit E(i). By the default, the deciles of E(i) are used to define 10 bins with equal frequencies of cases. The vertical axis can display either cumulative or noncumulative profit computed from C(i).
To see the effect on total profit of varying the decision threshold K, use a cumulative profit chart. Each percentile point p on the horizontal axis corresponds to a threshold K equal to the corresponding percentile of E(i). That is:
p/100 =[( Sum(i|E(i)<K) F(i))/(Sum(i) F(i))]
However, the chart shows only p, not K. Since the chart shows cumulative profit, each case with E(i) < K contributes a profit of C(i), while all other cases contribute a profit of zero. Hence, the ordinate (vertical coordinate) of the curve is the total profit for the decision rule Dk(i), assuming that the profit for the decision to do nothing is zero:
Sum(i|E(i)<K) F(i)C(i)
Transformations that add a constant t-hatto the tth row of the revenue matrix or a constant ci to the ith row of the cost matrix can change the expected profit for different cases by different amounts and therefore can alter the order of the cases along the horizontal axis of a profit chart, producing large changes in the cumulative profit curve.
To obtain a profit chart for the German credit data, you need to:
  1. Transform the decision matrix to have a column of zeros, as in the "Realistic Loss Matrix" above.
  2. Omit the zero column.
Hence, the decision matrix presented to the Model Comparison node should be:
Target Value
Decision
Accept
Good
–1
Bad
5
Copyright © SAS Institute Inc. All rights reserved.