Add methods for dealing with fairness in rankings #461

andrewklayk · 2023-09-08T10:42:27Z

We've implemented a set of post-processing algorithms for dealing with fairness in rankings, as described in https://arxiv.org/abs/1905.01989 (KDD 2019), complete with a demo notebook and tests.
Along with the algorithms, we've added two useful metrics for evaluating rankings that were introduced or referenced in the paper (Infeasibility Index and Discounted Cumulative Gain). Further details are available in the demo notebook.

closes #455

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

…ive and relaxed Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

…site Signed-off-by: Illia Kryvoviaz <illiakryvoviaz@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

hoffmansc

thanks @andrewklayk, this looks great! Other than the small comments, I think we can merge this.

I would love to include this in the aif360.sklearn subpackage as well. There is a slightly different API for those algorithms/metrics but it should actually be easier to use since you rely on DataFrames anyway. That can be a separate PR if you prefer or you can add it to this one.

examples/demo_deterministic_reranking.ipynb

hoffmansc · 2023-10-03T18:28:21Z

examples/demo_deterministic_reranking.ipynb

+    "import tempfile\n",
+    "import requests\n",
+    "import zipfile\n",
+    "import os\n",
+    "import pandas as pd \n",
+    "\n",
+    "with tempfile.TemporaryDirectory() as temp_dir:\n",
+    "    response = requests.get(\"http://www.seaphe.org/databases/LSAC/LSAC_SAS.zip\")\n",
+    "    temp_file_name = os.path.join(temp_dir, \"LSAC_SAS.zip\")\n",
+    "    with open(temp_file_name, \"wb\") as temp_file:\n",
+    "        temp_file.write(response.content)\n",
+    "    with zipfile.ZipFile(temp_file_name, 'r') as zip_ref:\n",
+    "        zip_ref.extractall(temp_dir)\n",
+    "    data = pd.read_sas(os.path.join(temp_dir, \"lsac.sas7bdat\"))\n",
+    "    data = data.assign(gender=(data[\"gender\"] == b\"male\") * 1)\n",
+    "    data['race'] = data['race1']\n",
+    "    data = data[['race', 'gender', 'lsat', 'ugpa', 'zfygpa']]\n"


can we not use the built-in LawSchoolGPADataset class here?

That was the initial idea, but I couldn't get the tempeh import to work, despite having installed the .[LawSchoolGPA] option of the package. I've tried installing tempeh separately, reinstalling AIF360 manually and so on, but to no end. Without the import, neither the LawSchoolGPADataset nor the fetch_lawschool_gpa work.

An update on this: it seems like the SEAPHE website is no more (domain expired), which means that neither this code nor the AIF360 functions work. Could you suggest some other dataset in the library that would work as an example here? Or, alternatively, the LawSchoolGPA dataset is still available from the Internet Archive backup of the SEAPHE website.

that is unfortunate... I've reached out to SEAPHE and they're investigating. Let's give it a week and see if they can get the website back up.

tests/test_demo_deterministic_reranking.py

aif360/algorithms/postprocessing/deterministic_reranking.py

aif360/metrics/regression_metric.py

aif360/metrics/__init__.py

aif360/algorithms/postprocessing/deterministic_reranking.py

examples/demo_deterministic_reranking.ipynb

aif360/algorithms/postprocessing/deterministic_reranking.py

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

andrewklayk · 2023-10-25T10:53:16Z

Hello @hoffmansc,
Thank you for your remarks. I've addressed the issues, except for the usage of the SEAPHE dataset. The site is still down, and I've been unable to find a free-use dataset from a reputable source that would fit here as an example.
In my opinion, the "toy example" in the notebook is enough to explain the API, so perhaps we should ditch the dataset part alltogether?
As for the aif360.sklearn - I think it's better to do this in a separate issue.

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

rahulnair23 · 2023-11-10T16:20:32Z

Hi @hoffmansc - seems like the Law School dependency has been removed from the demo notebook. Is this good to merge?

andrewklayk changed the title ~~Main rerank~~ Add methods for dealing with fairness in rankings Sep 8, 2023

andrewklayk and others added 20 commits September 8, 2023 13:00

added fairness-aware ranking

587a19c

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

Refactored reranking

e411e7c

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

Fixed indexing error in greedy

7e79d56

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

Minor refactorings to reranking

4f669ca

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

Minor refactorings to reranking

31df48c

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

added metrics and tests

02a9f4e

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

added rerank tests

780880a

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

added demo_deterministic_reranking.ipynb

a9dc191

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

added demo_deterministic_reranking tests

6282d5c

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

minor fixes to tests

56fb454

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

big rework of demo_deterministic_reranking, minor fixes

0dd0344

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

minor fixes to demo_deterministic_reranking

c64217d

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

minor fixes and refactorings

b3b08b4

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

work on allowing groups with several sensitive attributes

4d9052f

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

further work on multiple attributes, added tie-breakers for conservat…

e1628df

…ive and relaxed Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

updates to all deterministic_reranking algs

c614619

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

minor fixes in rerank example

51b6cb4

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

minor fix to constrained reranking and demo

ead4e82

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

adapted tests_deterministic_reranking.py to new signature

a5ee96c

Signed-off-by: andrewklayk <andrew.klyachkin@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

reworked the reranking example to take LawSchoolGPA directly from web…

c1aebb5

…site Signed-off-by: Illia Kryvoviaz <illiakryvoviaz@gmail.com> Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

andrewklayk force-pushed the main-rerank branch from a449dcd to c1aebb5 Compare September 8, 2023 11:03

fixed bug in dcg, slight changes to notebook

96b1825

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

andrewklayk force-pushed the main-rerank branch from eed62cd to 96b1825 Compare September 16, 2023 22:12

andrewklayk added 2 commits September 17, 2023 21:20

added theory to notebook

42c1d01

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

fixed bug in NDCG and reranking

c860145

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

andrewklayk force-pushed the main-rerank branch from 6fef28d to c860145 Compare September 17, 2023 19:53

andrewklayk added 2 commits September 17, 2023 22:41

fixed regressionmetric tests for infeasibleindex

ef13174

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

added small explainer to demo_deterministic_reranking

9bda6e7

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

hoffmansc requested changes Oct 3, 2023

View reviewed changes

hoffmansc reviewed Oct 3, 2023

View reviewed changes

aif360/algorithms/postprocessing/deterministic_reranking.py Outdated Show resolved Hide resolved

andrewklayk added 6 commits October 4, 2023 13:16

fixed typos in notebook

e164bde

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

removed incompatible type hints

6a8f085

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

removed test_demo_deterministic_reranking.py as unnecessary

93df903

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

added references

9ea7d7c

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

updated algorithms init file

c2be6f5

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

updated documentation with DeterministicReranking

5328843

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

andrewklayk added 3 commits October 25, 2023 16:45

fixed error in deterministic_reranking signature

8417296

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

bugfix, demo rework

73aea62

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

bugfix: >=1 sens.attr. support, error in Constrained alg

8ac9839

Signed-off-by: Andrii Kliachkin <andrew.klyachkin@gmail.com>

hoffmansc approved these changes Nov 10, 2023

View reviewed changes

hoffmansc merged commit 01b77d4 into Trusted-AI:master Nov 10, 2023
5 of 9 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add methods for dealing with fairness in rankings #461

Add methods for dealing with fairness in rankings #461

andrewklayk commented Sep 8, 2023

hoffmansc left a comment

hoffmansc Oct 3, 2023

andrewklayk Oct 4, 2023

andrewklayk Oct 5, 2023 •

edited

Loading

hoffmansc Oct 9, 2023

andrewklayk commented Oct 25, 2023 •

edited

Loading

rahulnair23 commented Nov 10, 2023

Add methods for dealing with fairness in rankings #461

Add methods for dealing with fairness in rankings #461

Conversation

andrewklayk commented Sep 8, 2023

hoffmansc left a comment

Choose a reason for hiding this comment

hoffmansc Oct 3, 2023

Choose a reason for hiding this comment

andrewklayk Oct 4, 2023

Choose a reason for hiding this comment

andrewklayk Oct 5, 2023 • edited Loading

Choose a reason for hiding this comment

hoffmansc Oct 9, 2023

Choose a reason for hiding this comment

andrewklayk commented Oct 25, 2023 • edited Loading

rahulnair23 commented Nov 10, 2023

andrewklayk Oct 5, 2023 •

edited

Loading

andrewklayk commented Oct 25, 2023 •

edited

Loading