Paper: arXiv 2306.10591

Authors: Gerhard Hellstern, Vanessa Dehn, Martin Zaefferer

Abstract

The problem of selecting an appropriate number of features in supervised learning problems is investigated in this paper. Starting with common methods in machine learning, we treat the feature selection task as a quadratic unconstrained optimization problem (QUBO), which can be tackled with classical numerical methods as well as within a quantum computing framework. We compare the different results in small-sized problem setups. According to the results of our study, whether the QUBO method outperforms other feature selection methods depends on the data set. In an extension to a larger data set with 27 features, we compare the convergence behavior of the QUBO methods via quantum computing with classical stochastic optimization methods. Due to persisting error rates, the classical stochastic optimization methods are still superior.

Complexity vs Empirical Score

  • Math Complexity: 8.0/10
  • Empirical Rigor: 4.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper introduces complex quantum computing frameworks (QUBO, QAOA, VQE) and advanced mathematical formulations, but its empirical evaluation is limited to small, simplified datasets with no mention of backtesting or real-world trading implementation.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Optimize Feature Selection for<br>Supervised Learning"] --> B["Formulate as QUBO<br>Quadratic Unconstrained Binary Optimization"]
  B --> C{"Problem Size?"}
  
  C --> D["Small Dataset<br>n < 27 features"]
  C --> E["Large Dataset<br>n = 27 features"]
  
  D --> F["Compare: Classical Numerical<br>vs Quantum Computing"]
  E --> G["Compare: Classical Stochastic<br>vs Quantum Computing"]
  
  F --> H["Outcome: Performance depends<br>on specific dataset"]
  G --> I["Outcome: Classical methods superior<br>due to quantum error rates"]
  
  H --> J["Conclusion:<br>Current quantum frameworks<br>limit competitive advantage"]
  I --> J

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes formulating feature selection as a quadratic unconstrained optimization problem (QUBO) and solving it with both classical and quantum computing methods.

Source sentence
we treat the feature selection task as a quadratic unconstrained optimization problem (QUBO), which can be tackled with classical numerical methods as well as within a quantum computing framework.

Assumes 5

  • Logistic regression is used as the machine learning model for all experiments. model
    Source sentence
    we use logistic regression as the machine learning model and use the whole data set for training.
  • The whole data set is used for training, which may lead to overfitting. statistical
    Source sentence
    We are aware, that this may lead to overfitting, however, logistic regression, in comparison to more sophisticated models, proves often to be robust against overfitting
  • Binary classification is the task for both data sets. data
    Source sentence
    we seek a binary classification model with the "optimal" number of features.
  • Dependency measures (correlation, mutual information, ROC, ANOVA) are assumed to capture feature relevance. model
    Source sentence
    In the following, for the inter-feature dependence we restrict ourselves to correlation and the mutual information measure. For the dependence on the target, we tested all measures.
Show 1 more
  • The weighting parameter ϕ balances intra‑feature dependence and feature‑target dependence. model
    Source sentence
    the weight for the two conditions (intra-feature dependence and dependence between features and target) depends on the choice of the dependency measure and it is not useful to be set globally

Data

  • Wisconsin breast cancer data other public
    Source sentence
    ... we have 10 features and 569 observations in the data set

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • The QUBO‑based optimization outperforms RFE and is comparable to LASSO, achieving the best AUROC when using ROC between features and correlation between features and target. performance
    Source sentence
    RFE is outperformed by our proposed optimization approach with a suitable choice of dependency measure. The LASSO result is only slightly worse than the best possible.
  • On the lending club data the quantum backend reproduces the optimal objective value obtained by classical optimizers. theoretical
    Source sentence
    Using the lending club data, we obtain the same optimal value for the objective function, obj = −3.7803, with the classical optimizer, the direct diagonalization of the Hamiltonian, the state vector simulator, and with the quantum simulator.
  • QAOA on actual quantum hardware achieves decent early results but does not yet match non‑quantum metaheuristics. negative
    Source sentence
    While QAOA (when run on actual QC hardware) achieves decent results fairly early, it is not yet able to compete with non‑QC optimization algorithms.
  • Higher values of the weighting parameter ϕ (≈0.95) generally improve AUROC for the lending club data. performance
    Source sentence
    In all of our experiments, we find the highest value of the AUROC when using the ROC as the dependency measure between the features and the correlation or mutual information between each feature and the target.
  • For the breast cancer data the best AUROC is achieved with the (MI, MI) combination at ϕ around 0.75. performance
    Source sentence
    we find the highest value of the AUROC for the combination (MI, MI), but only for a narrow band of ϕ-values around 0.75.

Checked by others

Supported by

  • Improved Financial Forecasting via Quantum Machine Learning (2023)
    On the question: Can quantum computing currently enhance feature selection or financial forecasting performance?
    Compare the findings

    This paper: QAOA on actual quantum hardware achieves decent early results but does not yet match non‑quantum metaheuristics.

    That paper: Current quantum hardware is not powerful enough to provide real improvements; performance gains can be achieved today by turning quantum ideas into classical ML solutions.

    Both findings agree that present quantum hardware does not yet outperform classical methods; A reports early decent QAOA results but falls short of metaheuristics, while B notes gains are only achievable by converting qu

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.