You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
<divclass="pub-group-heading">Beyond Data Structures</div>
222
222
223
+
<articleclass="paper-card" id="unbiased-binning">
224
+
<divclass="paper-tag">VLDB 2026</div>
225
+
<h3>Unbiased Binning</h3>
226
+
<pclass="paper-meta">
227
+
Abolfazl Asudeh, Zeinab Asoodeh, Bita Asoodeh, and Omid Asudeh · <em>Proceedings of the VLDB Endowment</em>, Vol. 19, 2026
228
+
</p>
229
+
<divclass="paper-feature">
230
+
<figureclass="paper-figure">
231
+
<imgsrc="imgs/UnbiasedBinning.jpg" alt="Illustration for the Unbiased Binning paper">
232
+
</figure>
233
+
234
+
<divclass="paper-feature-copy">
235
+
<p>
236
+
Discretizing raw features into bucketized attributes is a common step before sharing a dataset. However, this process can inadvertently introduce bias and amplify unfairness in downstream tasks.
237
+
</p>
238
+
<p>
239
+
This work formulates the unbiased binning problem, which seeks bucketized attributes that satisfy group parity. The paper develops an efficient dynamic programming algorithm for equal-size binning, while also showing that perfect parity may come with a high price of fairness or may not exist when group distributions differ substantially.
240
+
</p>
241
+
<p>
242
+
To support settings where small deviations from perfect parity are acceptable, the paper introduces <em>epsilon</em>-biased binning, which limits group disparities across buckets to at most <em>epsilon</em>. It first presents a quadratic-time dynamic programming algorithm, then addresses scalability with a local search strategy whose divide-and-conquer component finds valid solutions in near-linear time whenever one exists.
243
+
</p>
244
+
<p>
245
+
The local search and divide-and-conquer algorithms are general and are not limited to equal-size binning. Extensive experiments on real-world and synthetic datasets confirm the efficiency of the algorithms and show that fairness-unaware binning can generate biased attribute representations, while fairness-aware binning can significantly reduce this bias with negligible price of fairness.
Abolfazl Asudeh, Zeinab Asoodeh, Bita Asoodeh, and Omid Asudeh · <em>Proceedings of the VLDB Endowment</em>, Vol. 19, 2026
363
-
</p>
364
-
<divclass="paper-feature">
365
-
<figureclass="paper-figure">
366
-
<imgsrc="imgs/UnbiasedBinning.jpg" alt="Illustration for the Unbiased Binning paper">
367
-
</figure>
368
-
369
-
<divclass="paper-feature-copy">
370
-
<p>
371
-
Discretizing raw features into bucketized attributes is a common step before sharing a dataset. However, this process can inadvertently introduce bias and amplify unfairness in downstream tasks.
372
-
</p>
373
-
<p>
374
-
This work formulates the unbiased binning problem, which seeks bucketized attributes that satisfy group parity. The paper develops an efficient dynamic programming algorithm for equal-size binning, while also showing that perfect parity may come with a high price of fairness or may not exist when group distributions differ substantially.
375
-
</p>
376
-
<p>
377
-
To support settings where small deviations from perfect parity are acceptable, the paper introduces <em>epsilon</em>-biased binning, which limits group disparities across buckets to at most <em>epsilon</em>. It first presents a quadratic-time dynamic programming algorithm, then addresses scalability with a local search strategy whose divide-and-conquer component finds valid solutions in near-linear time whenever one exists.
378
-
</p>
379
-
<p>
380
-
The local search and divide-and-conquer algorithms are general and are not limited to equal-size binning. Extensive experiments on real-world and synthetic datasets confirm the efficiency of the algorithms and show that fairness-unaware binning can generate biased attribute representations, while fairness-aware binning can significantly reduce this bias with negligible price of fairness.
0 commit comments