Advanced techniques with Singularity Machine Learning - Classification

Use the following techniques to get the most out of Singularity Machine Learning - Classification. For every parameter and its default value, see the API reference.

---

Prototype on a simulator

Set optimizer_options={"simulator": True} in the create or create_fit_predict action to run the QAOA optimization on a simulator instead of a QPU. The examples in the get started guide use this setting. When your workflow is ready, remove it or set it to False to run on IBM® QPUs.

---

Size the ensemble

num_learners sets the size of the ensemble, which corresponds to the number of qubits used. Larger ensembles can lead to increased accuracy, as shown in the benchmarks. Use learners_types, learners_proportions, and learners_options to mix scikit-learn classifiers in the ensemble.

---

Tune regularization

• regularization_type is onsite or alpha. Higher onsite values and lower alpha values lead to sparser ensembles.
• Set regularization to auto to search for a regularization value that selects the ratio of learners given by regularization_desired_ratio.
• Pass a list to regularization_type or regularization to train a model for each value and keep the best performing one.

---

Use validation data

Pass fit_params={"validation_data": (X_val, y_val)} to the fit action, and set optimization_data to train, validation, or both to choose the data that the quantum optimization uses.

---

Choose a voting method

Set voting to hard for majority voting, or to soft to average the learners' probabilities. Use prob_threshold to set the probability threshold, which defaults to 0.5.

---

Make results repeatable

Set random_state to control randomness, so that you can compare runs with different settings.