We develop an interpretable and scalable statistical framework, which combines experimental data with a mechanistic model to infer transcriptional burst kinetics (sizes and frequencies) and feedback regulations. Applying this framework to the scRNA-seq data of embryonic mouse fibroblast cells, we find that genome-wide burst kinetics exhibit different characteristics in cases without and with feedback regulations, implying Simpson’s paradoxes. We show that feedbacks modulate burst frequencies and sizes differently and conceal the effects of transcription start site distributions on burst kinetics. Notably, only in the presence of positive feedback, TATA genes are expressed with high burst frequencies, and enhancer-promoter interactions mainly modulate burst frequencies.