Should HSPEstimator be a regressor, a classifier, or both?

[Gnpd]

Context

In HSPiPy, the HSPEstimator class is implemented as a scikit-learn–compatible estimator.
Currently, its role is somewhat ambiguous:

• Classifier perspective:
  predict(X) returns binary labels (1 if a solvent is inside the fitted sphere(s), 0 otherwise).
  Accuracy is measured using classification metrics.

• Regressor perspective:
  transform(X) returns continuous RED values (relative energy distance).
  These are closer to regression outputs, and one could imagine fitting/predicting continuous solubility metrics instead of hard inside/outside classification.

• Transformer perspective:
  It also works inside pipelines as a transformer, converting solvent coordinates into RED distances.

Binarization of solvent scores

At present, solvent labels are binarized inside the estimator:

((y <= inside_limit) & (y != 0)).astype(int)

This works well when datasets already classify solvents as “good/bad” (inside/outside).
However, many published datasets provide graded solubility values (e.g. swelling %, absorbance, weight loss, Hansen distances, etc.) rather than a simple binary score.

Possible implications:

• A classifier design enforces binarization, possibly losing information in graded datasets.
• A regressor design could predict continuous solubility scores directly.
• A hybrid/dual approach (e.g. HSPClassifier and HSPRegressor) might allow both.

Why this matters

• Some applications (polymer screening, formulation) only need good/bad classification.
• Others (quantitative solubility, miscibility ranges) would benefit from regression on the raw scores.
• Choosing "classifier" vs "regressor" affects which metrics (accuracy_score vs r2_score) are available and how pipelines integrate HSPiPy.

Questions for the community

1. Do you mainly work with binary good/bad solvent datasets, or with continuous solubility ranges?
2. Would splitting into HSPClassifier and HSPRegressor make sense, or should we keep a single flexible class?
3. Are there published workflows where regression on raw scores gives better HSP fits than classification on binarized labels?

---

💬 Feedback is very welcome! Please share your preferences, experiences, or links to relevant publications on handling continuous vs binary solubility data.