A cost matrix assigns a cost to each combination of predicted class and actual class. In cost-sensitive learning, these costs weight each instance's loss so costly mistakes matter more. section 5.2.2
For binary classification, entries distinguish true negatives, false negatives, false positives, and true positives using the convention (predicted, actual). A useful matrix requires a problem context where decision costs can be quantified. section 5.2.2
Virtual screening
A false negative for an inhibitor can waste downstream drug-discovery resources, giving that error a measurable cost. Investment decisions provide another example when false positives have known monetary consequences. section 5.2.2
Constructing the matrix requires extra computation and reliable cost estimates. When those estimates are unavailable, class-balanced loss offers a simpler alternative by weighting classes inversely to their sample counts. section 5.2.2
Table 5.1 Cost matrix for cost-sensitive learning. Subscripts read (predicted, actual)
|
Predicted ↓
/ Actual →
|
Actual Negative
|
Actual Positive
|
|
Predicted Negative
|
C00 (TN)
|
C01 (FN)
|
|
Predicted Positive
|
C10 (FP)
|
C11 (TP)
|
Use the binary matrix to inspect how predicted and actual classes map to TN, FN, FP, and TP costs. section 5.2.2

