In the Model Screening platform, the Decision Thresholds report enables you to explore thresholds for binary classification models. There is a Decision Thresholds report for each model data set that is specified by the validation method. These could be Training, Training and Validation, or Training, Validation, and Test sets. Each Decision Thresholds report contains a graph of the distribution of fitted probabilities, a bar chart of classifications, and confusion matrices. All are organized by model fit, fold, trial, and class level. The report also contains a tabbed section on classification accuracy measures and an option to set the profit matrix. The report updates as you adjust the probability threshold.
The distribution of fitted probabilities, or model scores, enables you to see how each individual model fit differentiates between the two classes. A vertical line on the graph represents the probability threshold, which determines the classification of each observation. By default, the probability threshold is 0.5. You can change the probability threshold by dragging the vertical line or by clicking the Probability Threshold value and entering a new value. This changes the probability threshold value across the whole Decision Thresholds report. The value of the probability threshold must be between 0 and 1.
The bar chart of classifications shows the classification counts for each level of the response variable at the current threshold. Green bars represent correctly classified observations; red bars represent incorrectly classified observations.
The confusion matrices, also known as contingency tables, show the two-way classification of actual and predicted responses for each individual model fit. Confusion rates matrices are shown as well. The rates are equal to the values in the confusion matrices divided by the row totals.
False Classification by Threshold
Shows a plot of the misclassification count by probability threshold and a plot of the misclassification rate by probability threshold. Each plot contains two curves for each individual model fit. The curves for the low response category are solid and the curves for the high response category are dashed. The curves intersect at the threshold that yields equal misclassification, counts or rates, of each response level. There is also a vertical line on each graph that represents the current probability threshold value. You can change the probability threshold value by dragging the vertical line. This changes the probability threshold value across the whole report.
False Classification by Portion
Shows a plot of the misclassification count or rate by the portion of the rank ordered scores. Each plot contains two curves for each individual model fit. The curves for the low response category are solid and the curves for the high response category are dashed.
True Classification by Threshold
Shows a plot of the true count by the probability threshold and a plot of the true rate by the probability threshold. Each plot contains two curves for each individual model fit. The curves for the low response category are dashed and those for the high response level are solid. The curves intersect at the threshold that yields equal correct classifications, counts or rates, for each response. There is also a vertical line on each graph that represents the current probability threshold value. You can change the probability threshold value by dragging the vertical line. This changes the probability threshold value across the whole report.
True Classification by Portion
Shows a plot of the true count or rate by the portion of the rank ordered scores. Each plot contains two curves for each individual model fit. The curves for the low response category are dashed and the curves for the high response category are solid.
Profit by Threshold
(Available only if a profit matrix is specified.)Shows a plot of the average profit by the probability threshold. There is a curve for each individual model fit and a vertical line that represents the current probability threshold value. The specified profit matrix is also shown next to the plot.
Metrics
Shows a table of classification accuracy metrics for each model. A legend is provided to describe how the metric in each column is calculated.
Note: Two less common classification accuracy metrics are F1 and MCC. The F1 score is a combination of precision and recall, or sensitivity. Another way to calculate F1 is 2(Precision × Sensitivity)/(Precision + Sensitivity). The Mathews correlation coefficient (MCC) is equivalent to the Pearson correlation coefficient estimated for two binary variables. See Statistical Details for the Pearson Product-Moment Correlation.
Enables you to assign costs to undesirable outcomes and profits to desirable outcomes. See Specify Profit Matrix. If you change the probability threshold in the profit matrix window and click OK, the Decision Thresholds report is updated using that value as the probability threshold.