Architecture

Qmes implements a two-stage meta-learning pipeline. The quantum-expensive work happens once, offline; a user at inference time runs only classical code.

Stage (a): data processing and ground-truth label generation. Stage (b): recommender training with multiple classifiers and feature subset configurations.

Core components

Three pluggable components, each an abstract base class with one concrete implementation per task type:

Component	Base class	Role	Phase
Extractor	BaseExtractor	Dataset \(\rightarrow\) fixed-length complexity meta-feature vector	offline and inference
Evaluator (Oracle)	BaseEvaluator	Performance score via quantum kernel \(\rightarrow\) meta-labels	offline only
Recommender	PairwiseRecommender	Meta-features \(\rightarrow\) ranked circuit list	offline fit, inference predict

Utility modules

Module	Contents
Preprocessing (Qmes.data.preprocessing)	Categorical encoding, median imputation, subsampling to 600 samples
Circuit registry (Qmes.circuits)	CIRCUIT_POOL, quantum kernel-matrix computation
Model selection (Qmes.recommender.selection)	LOO search over 14 classifiers × MI-selected feature subsets
Inference runner (Qmes.inference)	recommend, preprocess_new_dataset, evaluate_recommendation

The bundled Qsun simulator provides the quantum backend for all circuit evaluations and is invoked exclusively during the offline phase.

Extending Qmes

Both base classes use a small fixed contract, a subclass implements three members and the base class handles validation, sanitization, and the public API:

• New extractor: task_type, _feature_names, _extract_raw(X, y)
• New Oracle: task_type, metric_name, evaluate_circuit(X, y, name)
• New circuit: one dict entry in CIRCUIT_POOL - no subclassing

Worked, executed examples for all three are in Advanced Usage.