Cost-sensitive learning modifies model loss so different prediction errors receive different penalties. In virtual screening, a false negative may waste downstream drug-discovery resources, making error costs part of the modeling problem. section 5.2.2
A practical workflow includes:
- Define a cost matrix covering true negatives, false negatives, false positives, and true positives. section 5.2.2
- Use explicit costs when misclassification consequences can be quantified; otherwise, weight each class contribution to produce a class-balanced loss. section 5.2.2
- Compare imbalance strategies with precision-recall area under the curve rather than relying only on class 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)
|
Inspect the cost matrix that organizes error penalties by predicted and actual class. section 5.2.2

