In short
Yes: a Random Forest pre-stage made the VQC far better than PCA did, at every qubit count from 2 to 12 and on every one of the 50 splits. Two qubits were enough. The gain was largest where qubits are scarcest.
The primary research question asked if the performance of a VQC on a high-dimensional malware dataset could be improved using a classical machine learning model, instead of PCA, at the same qubit count. In an exact noiseless simulation, using the GBAC dataset, the answer was yes.
RQ1 – Does a Random Forest pre-encoding stage improve VQC accuracy compared to PCA for the same number of qubits?
Yes. At 4 qubits the leaf embedding arm beat PCA + VQC by +0.210 in accuracy, while the grouped votes arm beat it by +0.216, and at 8 qubits they beat PCA + VQC by +0.153 and +0.160 respectively. After the Bonferroni correction on all five metrics, the difference was significant and both RF arms beat PCA on all 50 splits.
RQ2 – Can a VQC with fewer qubits, using a Random Forest pre-encoding stage, match or outperform a PCA-based VQC with more qubits?
Yes. PCA + VQC peaked at an accuracy of 0.830 at 8 qubits, while at 2 qubits a VQC using an RF pre-stage got an accuracy of 0.992 for grouped votes and 0.986 for leaf embedding, and even at 12 qubits PCA + VQC still did not score better than either RF arm at 2 qubits. The best PCA split at any qubit count was 0.886, while the worst grouped votes split got 0.985.
RQ3 – How does the improvement change for VQCs with increasing numbers of qubits?
It did not grow with the number of qubits. For grouped votes the gap was largest at 2 qubits (+0.232), stayed between +0.217 and +0.223 for 3–6 qubits, narrowed to +0.160 to +0.162 at 8–10 qubits and was +0.170 at 12 qubits. The gap narrowed at 8 qubits because PCA improved, not because the RF arms got worse. The improvement was largest where qubits are scarcest, and it held at every qubit count tested.
Accuracy (%) by qubit count
| Method | 2 | 3 | 4 | 6 | 8 | 10 | 12* |
|---|---|---|---|---|---|---|---|
| RF votes + VQC | 99.2 | 99.2 | 99.2 | 99.1 | 99.1 | 99.1 | 99.0 |
| RF leaf + VQC | 98.6 | 98.5 | 98.5 | 98.5 | 98.4 | 98.4 | 98.3 |
| PCA + VQC | 76.0 | 77.5 | 76.9 | 76.9 | 83.0 | 82.8 | 82.0 |
*12 qubits: exploratory, 10 splits. For context, with no qubits: RF only 99.1, linear SVM 97.4, RBF SVM 83.3.
In detail
- RF pre-stage versus PCA at 4 and 8 qubits
- Performance across qubit counts (2–12)
- Comparison with classical models
- Ablation: what each pipeline change added
- Why it works