Figure 4

Comparison of classification strategies. Performance metrics for two distinct classification strategies: binary sequential (left column) and multiclass (right column). In the top panels, the confusion matrices for each strategy are shown; each collision is plotted as its predicted label (rows) and true label (columns). Predictions along the diagonal are correct classifications, whereas those in off-diagonal cells are misclassifications. In the middle panels, the distribution of masses resulting from false negatives (FN) and false positives (FP) are plotted. These mass residuals are important to constrain, because FP and FN predictions cannot be quantified in the regression stage. In the bottom panels, the classifier predictions are plotted along the \(b_{\infty }\)–\(v_{\infty }\) hyperplane, where gray indicates class 0 (no remnants), blue is class 1 (one remnant), orange is class 2 (two remnants), and misclassified collisions are indicated by red markers. The misclassified collisions are clustered near the transitions between classes