Feature Hashing

Hashes an input feature space to an n-bit feature space. Feature hashing is an efficient way of vectorizing features, and performing dimensionality reduction or expansion along the way. Supported types include array.array, list, dict, float, int, and string. The behaviour for different input data column types is as follows:

• array.array : The index of each element is combined with the column name and hashed, and the element becomes the value.

• list : Behaves the same as array.array, but if the element is non-numerical the element is combined with the column name and hashed, and 1 is used as the value.

• dict : Each key in the dictionary is combined with the column name and hashed, and the value is kept. If the value is is non-numerical, the element is combined with the column name and hashed, and 1 is used as the value.

• float : the column name is hashed, and the column entry becomes the value

• int : Same behavior as float

• string : Hash the string and use it as a key, and use 1 as the value.

The hashed values are collapsed into a single sparse representation of a vector. The number of bits to hash to is specified by the num_bits parameter.

Note: Each an entry is hashed, a seperate hash on the key is performed to either add or subtract a value with equal probability. This keeps the value unbiased, since the expectation value for each feature(across all examples) is 0.

Introductory Example

from graphlab.toolkits.feature_engineering import *

# Construct a feature hasher with default options.
sf = graphlab.SFrame({'a': [1,2,3], 'b' : [2,3,4], 'c': [9,10,11]})
hasher = graphlab.feature_engineering.create(sf, FeatureHasher())

# Transform the data using the hasher.
hashed_sf = hasher.transform(sf)

# Save the transformer.
hasher.save('save-path')

# Hash only a single column 'a'.
hasher = graphlab.feature_engineering.create(sf,
                                              FeatureHasher(features = ['a']))

# Hash all columns except 'a'.
hasher = graphlab.feature_engineering.create(sf,
                                    FeatureHasher(excluded_features = ['a']))

Fitting and transforming

Once a FeatureHasher object is constructed, it must first be fitted and then the transform function can be called to generate hashed features.

For numeric columns:

sf = graphlab.SFrame({'a' : [1,2,3], 'b' : [2,3,4]})
hasher = graphlab.feature_engineering.FeatureHasher()
fit_hasher = hasher.fit(sf)
hashed_sf = fit_hasher.transform(sf)

Columns:
        hashed_features dict

Rows: 3

Data:
+-------------------------+
|     hashed_features     |
+-------------------------+
| {79785: -1, 188475: -2} |
| {79785: -2, 188475: -3} |
| {79785: -3, 188475: -4} |
+-------------------------+
[3 rows x 1 columns]

For list/vector columns:

l1 = [1,2,3]
l2 = [2,3,4]
sf = graphlab.SFrame({'a' : [l1,l1,l1], 'b' : [l2,l2,l2]})
hasher = graphlab.feature_engineering.FeatureHasher()
fit_hasher = hasher.fit(sf)
hashed_sf = fit_hasher.transform(sf)

Columns:
    hashed_features dict

Rows: 3

Data:
+-------------------------------+
|        hashed_features        |
+-------------------------------+
| {642: 2.0, 164: -3.0, 937:... |
| {642: 2.0, 164: -3.0, 937:... |
| {642: 2.0, 164: -3.0, 937:... |
+-------------------------------+
[3 rows x 1 columns]

For string columns:

sf = graphlab.SFrame({'a' : ['a','b','c'], 'b' : ['d','e','f']})
hasher = graphlab.feature_engineering.FeatureHasher()
fit_hasher = hasher.fit(sf)
hashed_sf = fit_hasher.transform(sf)

Columns:
        hashed_features dict

Rows: 3

Data:
+------------------+
| hashed_features  |
+------------------+
| {405: 1, 79: 1}  |
| {454: 1, 423: 1} |
| {308: 1, 36: 1}  |
+------------------+
[3 rows x 1 columns]

For dictionary columns:

dict1 = {'a' : 1 , 'b' : 2 , 'c' : 3}
dict2 = {'d' : 4 , 'e' : 5 , 'f' : 6}
sf = graphlab.SFrame({'a' : [dict1, dict1, dict1],
                          'b' : [dict2, dict2, dict2]})
hasher = graphlab.feature_engineering.FeatureHasher()
fit_hasher = hasher.fit(sf)
hashed_sf = fit_hasher.transform(sf)

Columns:
    hashed_features dict

Rows: 3

Data:
+-------------------------------+
|        hashed_features        |
+-------------------------------+
| {36: 3, 454: 5, 423: 2, 79... |
| {36: 3, 454: 5, 423: 2, 79... |
| {36: 3, 454: 5, 423: 2, 79... |
+-------------------------------+
[3 rows x 1 columns]