|
|
 |
| |
|
|
|
Algorithmic Ambiguity and Ethical Decision Boundaries in News Recommendation Systems |
|
|
|
PP: 819-845 |
|
|
doi:10.18576/jsap/140622
|
|
|
|
Author(s) |
|
|
|
Yogeesh N,
Suleiman Ibrahim Mohammad,
N Raja,
Nosir Khurramov,
Asokan Vasudevan,
Ruzmetova Zilolakhon Dilmuradovna,
|
|
|
|
Abstract |
|
|
| Personalization in news delivery often amplifies user engagement but can introduce uncertain recommendations
and unintended bias. This study presents a unified mathematical framework that simultaneously quantifies recommendation uncertainty and enforces ethical exposure parity. We first introduce two complementary ambiguity metrics a threshold‐based score capturing local confidence and an information‐theoretic entropy measure for global uncertainty and model recommendation confidence via continuous membership functions. Ethical constraints are formalized as exposure‐gap thresholds between content groups, integrated into a matrix‐factorization objective alongside relevance and ambiguity penalties. An efficient alternating‐minimization algorithm, augmented with projection or smooth penalties, optimizes this composite loss. To illustrate practical behavior, we conduct a detailed synthetic case study on a 10×50 user–item dataset, computing per‐pair ambiguity, group exposures, fairness penalties, and composite scores. Quantitative results show that controlling uncertainty substantially reshapes recommendation rankings, while enforcing exposure‐parity incurs only a modest additional cost when tolerance levels align with observed disparity. Sensitivity analyses reveal how decision thresholds, penalty weights, and fairness tolerances influence the trade‐off surface. Our findings demonstrate that embedding rigorous uncertainty and fairness controls can yield news feeds that maintain high relevance, improve transparency, and respect ethical guidelines with minimal impact on user‐perceived quality. This framework offers practitioners a principled
pathway to design recommendation systems that balance personalization objectives with societal values. |
|
|
|
|
|
|
|
