We address the problem of efficiently computing Shap explanation scores for classifications with machine learning models. With this goal, we show the transformation of binary neural networks (BNNs) for classification into deterministic and decomposable Boolean circuits, for which knowledge compilation techniques are used. The resulting circuit is treated as an open-box model, to compute Shap scores by means of a recent efficient algorithm for this class of circuits. Detailed experiments show a considerable gain in performance in comparison with computing Shap directly on the BNN treated as a black-box model.