fairsearchdeltr 1.0.2

A Python library for disparate exposure in ranking (a learning to rank approach)

Fair search DELTR for Python

This is the Python library that implements the DELTR model for fair ranking.

Installation

To install fairsearchdeltr, simply use pip (or pipenv):

pip install fairsearchdeltr

And, that's it!

Using it in your code

You need to import the class from the package first:

from fairsearchdeltr import Deltr

Train a model

You need to train the model before it can rank documents.

# import other helper libraries
import pandas as pd
from io import StringIO

# load some train data (this is just a sample - more is better)
train_data_raw = """q_id,doc_id,gender,score,judgment
    1,1,1,0.962650646167003,1
    1,2,0,0.940172822166108,0.98
    1,3,0,0.925288002880488,0.96
    1,4,1,0.896143226020877,0.94
    1,5,0,0.89180775633204,0.92
    1,6,0,0.838704766545679,0.9
    """
train_data = pd.read_csv(StringIO(train_data_raw))

# setup the DELTR object
protected_feature = "gender" # column name of the protected attribute (index after query and document id)
gamma = 1 # value of the gamma parameter
number_of_iterations = 10000 # number of iterations the training should run
standardize = True # let's apply standardization to the features

# create the Deltr object
dtr = Deltr(protected_feature, gamma, number_of_iterations, standardize=standardize)

# train the model
dtr.train(train_data)
>> array([0.02527054, 0.07692437])
# your run should have approximately same results  

Use the model to rank

Now, you can use the obtained model to rank some data.

# load some test/prediction data
prediction_data_raw = """q_id,doc_id,gender,score
    1,7,0,0.9645
    1,8,0,0.9524
    1,9,0,0.9285
    1,10,0,0.8961
    1,11,1,0.8911
    1,12,1,0.8312
    """
prediction_data = pd.read_csv(StringIO(prediction_data_raw))

# use the model to rank the data  
dtr.rank(prediction_data)
>> doc_id  gender  judgement
4      11       1   0.074849
5      12       1   0.063770
0       7       0   0.063486
1       8       0   0.061248
2       9       0   0.056828
3      10       0   0.050836
# the result will be a re-ranked dataframe

The library contains sufficient code documentation for each of the functions.

Checking the model a bit deeper

You can check how the training of the model progressed using a special property called log.

dtr.log
>> [<TrainStep [1553844278383,[0.01926469 0.00976336],[[-0.00125304 -0.0014605 ]
  [-0.00125304 -0.0014605 ]
  [-0.00125304 -0.0014605 ]
  [-0.00125304 -0.0014605 ]
  [-0.00125304 -0.0014605 ]
  [-0.00125304 -0.0014605 ]],5.999620187652397,0.0]>,
 ...]

The log returns a list of objects from the fairsearchdeltr.models.TrainStep class. The class is a representation of the parameters in each step of the training. Contains a timestamp, omega, omega_gradient, loss, loss_standard, loss_exposure.

Development

1. Clone this repository git clone https://github.com/fair-search/fairsearchdeltr-python
2. Change directory to the directory where you cloned the repository cd WHERE_ITS_DOWNLOADED/fairsearchdeltr-python
3. Use any IDE to work with the code

Testing

Just run:

python setup.py test 

Credits

The DELTR algorithm is described in this paper:

• Meike Zehlike, Gina-Theresa Diehn, Carlos Castillo. "Reducing Disparate Exposure in Ranking: A Learning to Rank Approach." preprint arXiv:1805.08716 (2018).

This library was developed by Ivan Kitanovski based on the paper. See the license file for more information.

For any questions contact Mieke Zehlike

See also

You can also see the DELTR for ElasticSearch and DELTR Java library.