FeatureHasher¶

Implements the hashing trick.

Each pair of (name, value) features is hashed into a random integer in [0, n_features), using the signed 32-bit MurmurHash3 of the feature's token. String values are hashed as "name=value" tokens and contribute 1; numeric values are hashed under "name" and contribute the value itself.

The hashing is performed in Rust, so the whole transform of an example happens in a single native call.

Parameters¶

• n_features

  Default → 1048576

  The number by which each hash will be moduloed by.

• seed

  Type → int | None

  Default → None

  Set the seed to produce identical results. When None, a random seed is drawn, so two instances will hash features to different buckets.

• alternate_sign

  Type → bool

  Default → True

  When True (the default), the sign bit of the hash is used to negate half of the contributions. This keeps the expected value of each bucket at zero, so hash collisions between unrelated features tend to cancel out rather than accumulate, which is especially helpful for small n_features. This matches scikit-learn's FeatureHasher.

Examples¶

import river

hasher = river.preprocessing.FeatureHasher(n_features=10, seed=42)

X = [
    {'dog': 1, 'cat': 2, 'elephant': 4},
    {'dog': 2, 'run': 5}
]
for x in X:
    print(hasher.transform_one(x))

{5: -3, 7: 2}
{5: 2, 9: -5}

Methods¶

References¶

• Wikipedia article on feature vectorization using the hashing trick

• Weinberger et al. (2009), Feature Hashing for Large Scale Multitask Learning