Figure 7

SHAP results. (A) Top six features with highest SHAP scores for all models. SHAP scores with the largest average absolute value across all tested frames are shown for each behavior classifier. Each point denotes the SHAP score for one of the 150 frames included in SHAP analysis. The solid black line shows the base rate for each classifier, or the probability that the behavior occurs given the frequency of behavior in the training data set before consideration of feature information. Base rates vary by classifier and are shown in Supplemental Table S2. SHAP values greater than the base rate indicates positive SHAP scores for that frame, increasing the behavior probability. SHAP scores to the left of the black line indicate scores that reduce the behavior probability in that frame. Each point is colored by the z-scored value of the feature for that specific frame, indicating the relationship between feature values and shift in behavior probabilities. (B) Sum of SHAP scores of all features in each feature category for all classifiers. Positive SHAP scores indicate the magnitude of the feature influence on increasing the behavior probability when the behavior is present and negative SHAP values indicate influence on reducing the behavior probability in frames where the behavior is absent. 0 represents the base rate of behavior in the training dataset.