Add your own method¶

A transfer method receives TransferInputs, declares what fields it needs, and predicts scores for cold items.

Minimal method¶

import numpy as np
import pandas as pd

from warmtransfer.columns import Columns as C
from warmtransfer.methods.base import ColdStartMethod, register_method
from warmtransfer.types import TransferInputs


@register_method("constant_zero")
class ConstantZero(ColdStartMethod):
    requires = frozenset()

    def _fit(self, inputs: TransferInputs, seed: int) -> None:
        return None

    def predict(self, user_ids: np.ndarray, cold_item_ids: np.ndarray) -> pd.DataFrame:
        self._check_fitted()
        return pd.DataFrame(
            {
                C.User: np.repeat(user_ids, len(cold_item_ids)),
                C.Item: np.tile(cold_item_ids, len(user_ids)),
                C.Score: 0.0,
            }
        )

Declare inputs¶

Set requires to the fields your method actually needs:

donor_scores
content
similarity
embeddings
train_interactions
item_meta
val

The base class validates them before _fit runs.

Make it discoverable¶

Import the module from warmtransfer.methods.__init__ so @register_method(...) runs. Then add a unit test that fits the method on a tiny TransferInputs fixture and verifies the output columns and shape.

Use it in benchmark config¶

methods:
  - name: constant_zero