Table 2:

Results showing comparison of accuracy and macro-F1 scores across different interpretable models’ performance averaged over providers for the neutral-high classification. LR, DTC, SVM, RF and GBDT stand for Logistic Regression, Decision Tree Classifier, Support Vector Machine, Random Forest and Gradient Boosted Decision Trees respectively. Despite the small sample size and class imbalance, most models perform better than a macro-F1 threshold of 0.5

Social Signal	Metric	LR	DTC	SVM (Linear)	SVM (Radial)	RF	GBDT
Provider Dominance	Accuracy	0.719	0.673	0.703	0.704	0.722	0.717
	F1 Score	0.658	0.606	0.650	0.614	0.651	0.657
Provider Interactiveness	Accuracy	0.660	0.641	0.657	0.690	0.692	0.616
	F1 Score	0.559	0.550	0.548	0.519	0.523	0.525
Provider Engagement	Accuracy	0.863	0.947	0.852	0.903	0.898	0.934
	F1 Score	0.547	0.736	0.566	0.605	0.651	0.672
Provider Warmth	F1 Score	0.556	0.543	0.581	0.581	0.571	0.565
	F1 Score	0.515	0.501	0.544	0.460	0.521	0.515
Patient Dominance	Accuracy	0.975	0.968	0.973	0.973	0.961	0.970
	F1 Score	0.910	0.841	0.892	0.893	0.840	0.842
Patient Interactiveness	Accuracy	0.651	0.641	0.660	0.679	0.647	0.653
	F1 Score	0.559	0.509	0.557	0.525	0.523	0.519
Patient Engagement	Accuracy	0.826	0.663	0.827	0.883	0.804	0.826
	F1 Score	0.556	0.529	0.551	0.599	0.557	0.594
Patient Warmth	Accuracy	0.517	0.544	0.551	0.634	0.610	0.605
	F1 Score	0.537	0.562	0.568	0.604	0.614	0.603