In short
The VQC with a Random Forest in front scores about the same as the Random Forest on its own. So the win comes from replacing PCA, not from the quantum part. Both RF routes still beat both SVMs on every split.
Three classical models were used for context: RF only, a linear SVM and an RBF SVM (see baselines). They are not the bar to beat; the question remains RF + VQC against PCA + VQC. Their purpose is to show how much of the result comes from the forest itself, and where current classical methods sit. The accuracies were 0.991 for RF by itself, 0.974 for the linear SVM and 0.833 for the RBF SVM.
Per-fold accuracy of grouped votes (4 qubits) against RF only, linear SVM, RBF
SVM and PCA + VQC (8 qubits).
Votes at 4 qubits against RF only differ by only +0.0007 in accuracy, with votes winning 37, tying 4 and losing 9 of the 50 splits. This was statistically significant but far too small to matter in practice. At 8 qubits the difference was −0.0002 (20 wins, 4 ties, 26 losses), which is not significant and as such a tie. The leaf arm was 0.006–0.007 below RF only and lost all 50 splits, likely because it passes through SVD and loses a little in that process. A VQC of only 2 to 4 qubits therefore matches the accuracy of the full forest: the information the forest extracts is kept through the quantum stage, although the VQC does not improve on it.
Share of folds in which the row condition beats the column condition. Ties
count as half.
Both RF arms beat both SVMs on all 50 splits, by +0.010 to +0.017 against the linear SVM and +0.15 to +0.16 against the RBF SVM. PCA + VQC at 8 qubits ties the RBF SVM (−0.002, 25 wins and 25 losses). PCA + VQC at 4 qubits lost to all three classical models on all 50 splits.
Mean precision against recall per condition; circle size grows with qubit
count.
The RBF SVM misses about a third of the malware, with a precision of 0.988 but a recall of only 0.653. Its AUC of 0.983 means it ranks samples well but its decision threshold is poorly placed. The linear SVM is more balanced (precision 0.969, recall 0.977). Both RF arms sit top right, with the votes arm favouring precision (0.995) and the leaf arm having the highest recall of any condition (0.993–0.995). Since missing malware is the costly error here, leaf’s high recall is its main strength.
No quantum advantage
No quantum advantage is claimed, as the VQC does not beat the forest it receives its features from. A 2-qubit or 4-qubit VQC fed by the RF at best matches the best classical model here, while PCA + VQC does no better than the RBF SVM at its best.
Next: ablation