Evolution and impact of bias in human and machine learning algorithm interaction.

Area of Science:

• Computer Science
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Machine learning algorithms traditionally require expert labels but increasingly use public data, leading to potential bias.
• Algorithmic bias arises from unchecked data and interactive processes where both humans and algorithms receive biased information.
• Existing research often treats algorithmic bias as static, neglecting its dynamic, iterative nature in human-algorithm interactions.

Purpose of the Study:

• To investigate the dynamic and iterative nature of algorithmic bias.
• To introduce an iterated-learning framework inspired by human language evolution to model human-algorithm bias interaction.
• To analyze the impact of human action and iterated algorithmic bias modes (personalization filter, active learning, random) on machine learning performance.

Main Methods:

• Developed an iterated-learning framework to simulate human-algorithm interaction.
• Investigated three iterated algorithmic bias modes: personalization filter, active learning, and random.
• Formulated research questions to assess the impact of each bias mode.
• Conducted statistical analyses on results from controlled experiments.

Main Results:

• Iterated bias modes, initial data imbalance, and human action significantly affect machine learning model performance.
• Iterated filter bias, common in personalized interfaces, increases inequality and hinders human data discovery.
• A content-based filter created a 4% relevance blind spot; a real-life dataset simulation showed a 75% blind spot.

Conclusions:

• Algorithmic bias is a dynamic, iterative process influenced by human actions and algorithm selection mechanisms.
• Personalization filters can exacerbate bias, leading to reduced data relevance and discovery.
• The iterated-learning framework provides a valuable tool for understanding and mitigating long-term algorithmic bias effects.