Captured source text (L5)
6. Technical glossary Metrics used to measure the accuracy of age assurance
| Term | | | Meaning | | |
| --- | --- | --- | --- | --- | --- |
| Absolute error (AE) | | | | The same as the ‘error,’ but disregards the sign (i.e., positive | |
| | | | | or negative) thus focusing only on the magnitude (size) of the | |
| | | | | difference between the technologically-determined age and | |
| | | | | actual age. | |
| | Accuracy (ACC) | | The fraction of the predictions the model got right. The formula is ACC = (TP + TN) / (TP+ TN + FP + FN). | | |
| | Cumulative score (CS) | | | An aggregated score that is calculated by summing the | |
| | | | | individual score across over a period of time/category etc. | |
| Error | | | The user’s age determined by the technology minus the user’s actual age. An overestimation yields a positive value, whereas an underestimation yields a negative value. | | |
| False negative (FN) | | | | An outcome where a model incorrectly predicts a negative | |
| | | | | class i.e., a user is under 18 and the model predicts their age | |
| | | | | 18 or over. | |
| False negative rate (FNR) / Miss rate | | | Measures the proportion of FN against all negative predictions (i.e., FN and TP). FPR highlights the performance of the model in yielding FP results and this should be minimised. The formula is FNR = FN / (FN + TP). | | |
| False positives (FP) | | | | For the purpose of age assurance, this refers to an outcome | |
| | | | | where a model incorrectly predicts a positive class i.e., a user | |
| | | | | is 18 or over and the model predicts their age as under 18. | |
| False positive rate (FPR) | | | Measures the proportion of FP against all positive predictions (i.e., FP and TN). FPR highlights the performance of the model in yielding FP results and this should be minimised. The formula is FPR = FP / (FP + TN). | | |
| Mean absolute error (MAE) | | | | The central value of the absolute error. It describes the | |
| | | | | average discrepancy between a user's technology determined | |
| | | | | age and their actual age, ignoring whether it is an over- or | |
| | | | | under-estimation. It is calculated by summing the absolute | |
| | | | | errors for a given number of absolute errors, then dividing | |
| | | | | this by the number of absolute errors. The formula is MAE = | |
| | | | | (1/n) Σ(i=1 to n) \|y – x\| where n = number of observations in | |
| | | | | the dataset, y = is the true value, x = is the predicted value. | |
| Mean absolute percentage error (MAPE) | | | A metric that used to measure the accuracy in a regression analysis, this is useful where relative errors (age range estimations) are more meaningful than absolute errors. M = (1/n) Σ(t=1 to n)\|(At – Ft) / At) \|* 100 Where n = number of times the summation iteration happens, At = actual value and Ft = forecast value. | | |
Absolute error (AE) difference between the technologically-determined age and
Accuracy (ACC) formula is ACC = (TP + TN) / (TP+ TN + FP + FN).
Cumulative score (CS) individual score across over a period of time/category etc.
False positive rate (FPR) in yielding FP results and this should be minimised. The
Mean absolute error (MAE) errors for a given number of absolute errors, then dividing
25
| False negative rate (FNR) / Miss |
| --- |
| rate |
| Mean absolute percentage error |
| --- |
| (MAPE) |
| | Term | | | Meaning | |
| --- | --- | --- | --- | --- | --- |
| Access controls | Access controls | | | | |
| Age assurance | | | | | |
| Age assurance method | | | | | |
| Age assurance process | | | | | |
Full captured verbatim text
| Term | | | Meaning | | | | --- | --- | --- | --- | --- | --- | | Absolute error (AE) | | | | The same as the ‘error,’ but disregards the sign (i.e., positive | | | | | | | or negative) thus focusing only on the magnitude (size) of the | | | | | | | difference between the technologically-determined age and | | | | | | | actual age. | | | | Accuracy (ACC) | | The fraction of the predictions the model got right. The formula is ACC = (TP + TN) / (TP+ TN + FP + FN). | | | | | Cumulative score (CS) | | | An aggregated score that is calculated by summing the | | | | | | | individual score across over a period of time/category etc. | | | Error | | | The user’s age determined by the technology minus the user’s actual age. An overestimation yields a positive value, whereas an underestimation yields a negative value. | | | | False negative (FN) | | | | An outcome where a model incorrectly predicts a negative | | | | | | | class i.e., a user is under 18 and the model predicts their age | | | | | | | 18 or over. | | | False negative rate (FNR) / Miss rate | | | Measures the proportion of FN against all negative predictions (i.e., FN and TP). FPR highlights the performance of the model in yielding FP results and this should be minimised. The formula is FNR = FN / (FN + TP). | | | | False positives (FP) | | | | For the purpose of age assurance, this refers to an outcome | | | | | | | where a model incorrectly predicts a positive class i.e., a user | | | | | | | is 18 or over and the model predicts their age as under 18. | | | False positive rate (FPR) | | | Measures the proportion of FP against all positive predictions (i.e., FP and TN). FPR highlights the performance of the model in yielding FP results and this should be minimised. The formula is FPR = FP / (FP + TN). | | | | Mean absolute error (MAE) | | | | The central value of the absolute error. It describes the | | | | | | | average discrepancy between a user's technology determined | | | | | | | age and their actual age, ignoring whether it is an over- or | | | | | | | under-estimation. It is calculated by summing the absolute | | | | | | | errors for a given number of absolute errors, then dividing | | | | | | | this by the number of absolute errors. The formula is MAE = | | | | | | | (1/n) Σ(i=1 to n) \|y – x\| where n = number of observations in | | | | | | | the dataset, y = is the true value, x = is the predicted value. | | | Mean absolute percentage error (MAPE) | | | A metric that used to measure the accuracy in a regression analysis, this is useful where relative errors (age range estimations) are more meaningful than absolute errors. M = (1/n) Σ(t=1 to n)\|(At – Ft) / At) \|* 100 Where n = number of times the summation iteration happens, At = actual value and Ft = forecast value. | | | Absolute error (AE) difference between the technologically-determined age and Accuracy (ACC) formula is ACC = (TP + TN) / (TP+ TN + FP + FN). Cumulative score (CS) individual score across over a period of time/category etc. False positive rate (FPR) in yielding FP results and this should be minimised. The Mean absolute error (MAE) errors for a given number of absolute errors, then dividing 25 | False negative rate (FNR) / Miss | | --- | | rate | | Mean absolute percentage error | | --- | | (MAPE) | | | Term | | | Meaning | | | --- | --- | --- | --- | --- | --- | | Access controls | Access controls | | | | | | Age assurance | | | | | | | Age assurance method | | | | | | | Age assurance process | | | | | |