Classification with Variational Quantum Classifier
This is the repository of the group project of [4810-1166 Introduction to Quantum Computation / 量子計算論] in the University of Tokyo (U-Tokyo).
We performed classification on the Breast-Cancer dataset and predicted the survival of the Titanic passengers, whose dataset is provided at kaggle.
To reduce the number of involved qubits, we implemented a (3,1)-QRAC to encode 3 bits category feature into 1 qubit. In this way, we can even obtain a performance gain.
After encoding, we fed these encoded features to a VQC to perform classification.
- We trained a random forest classifier on Breast-Cancer dataset.
and selected four features according to the feature importance as
"tumor-size" "node-caps", "deg-malig", "menopause". - Then we encoded these category features with qubits, using QRAC.
- Besides, we built the corresponding circuit as FeatureMap.
- Finally, we trained a VQC with QRAC FeatureMap over all of the data. All of the parameters were optimized by SPSA optimizer with 100 iterations. And our model was evaluated on test set.
- We oversampled (doubled) the postive samples to balance the dataset.
Before the training step, we did some preprocessing jobs on Titanic dataset.
- For simplicity, we removed
"PassengerId", "Name", "Ticket", "Cabin"features because they are not so related to the survival or hard to process. - Since there are some values missing in the Titanic dataset,
firstly we masked them as
<NULL>and numericalized the rest of existing values. As for"Sex", "Embarked", we numericalized them into ordinal features. As for others, we just kept their original values since they are integer or float numbers. - Secondly, we filled the empty values according to some rules.
As for
"Sex", "Embarked", we randomly gave them values according to the distribution of the existing data. As for the rest of features, we fixed the missing values with the average of the existing data.
- Similarly, we also trained a random forest classifier on Titanic dataset.
and selected four features according to the feature importance as
"Sex", "Age", "Pclass", "Fare". - Then we encoded these category features with qubits, using QRAC.
- Besides, we built the corresponding circuit as FeatureMap.
- Finally, we trained a VQC with QRAC FeatureMap over all of the data. All of the parameters were optimized by SPSA optimizer with 100 iterations. And our model was evaluated on test set.
| Dataset | Breast-Cancer | Titanic | ||||||
|---|---|---|---|---|---|---|---|---|
| Model | Train Acc | Train F1 | Test Acc | Test F1 | Train Acc | Train F1 | Test Acc | Test F1 |
| :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: | :-: |
| Random Forest | 0.9474 | 0.9104 | 0.7759 | 0.3810 | 0.9798 | 0.9734 | 0.7344 | -/- |
| VQC | 0.6667 | 0.5190 | 0.5690 | 0.3902 | 0.6734 | 0.5727 | 0.6292 | -/- |
| VQC (real) | -/- | -/- | -/- | -/- | -/- | -/- | -/- | -/- |
| VQC + QRAC | 0.7325 | 0.5674 | 0.7241 | 0.4667 | 0.7868 | 0.7104 | 0.7656 | -/- |
| VQC + QRAC (real) | 0.7325 | 0.5612 | 0.7069 | 0.4849 | -/- | -/- | -/- | -/- |
# VQC
python3 main.py --encoder --featmap ZZMap --optimizer SPSA --backend simulator --dataset titanic
# VQC + QRAC
python3 main.py --encoder --featmap QRAC --optimizer SPSA --backend simulator --dataset titanic- Python 3.7
- qiskit 0.19.2
- numpy 1.18.5
- sklearn 0.23.1
[1] Yano, H., Suzuki, Y., Raymond, R., & Yamamoto, N. (2020). Efficient Discrete Feature Encoding for Variational Quantum Classifier. arXiv preprint arXiv:2005.14382.
- @Zhihong stated that we can fix the missing values of
"Sex"according to the name ("Mr.", "Ms.", etc.). - We plan to evaluate our model on the titanic dataset on real devices.