演化博弈视角下的算法偏见与伦理规制

Algorithmic Bias and Ethical Regulation under the Perspective of Evolutionary Game

推荐算法已成为社交媒体时代信息分发的驱动要素,受政治、资本、流量等因素影响,算法在信息分发过程中往往会出现结构性偏见、选择性偏见和趋利性偏见。本文基于人机演化博弈视角,对冷启动、协同过滤、收敛稳定和交互博弈四大阶段算法偏见的生成与纠偏机制进行分析,并基于反向工程实验,面向今日头条新闻客户端进行动态数据追踪和多阶段对比分析。发现算法推荐主题内容存在从发散到收敛的动态演变,使用早期旨在拓展用户兴趣边界,后期推荐主题趋于集中;在各个阶段正面内容均占据推荐主导地位;随着用户使用频次增加,中央媒体内容和广告内容被推荐占比趋于上升;面向不同群体推荐主题内容和情感倾向存在一定结构性差异。适度约束算法权力,有效保障用户权利,才能在合乎伦理的前提下最大化发挥人工智能的信息红利。

Recommendation algorithm has become the driving factor of information distribution in the era of social media.Influenced by political,capital,flow and other factors,the algorithm often presents structural bias,selective bias and profit bias in the process of information distribution.Based on the perspective of human-computer evolutionary game,this paper analyzes the generation and correction mechanism of algorithm bias in four stages:cold-start,collaborative filtering,convergence stability and interactive game.Based on reverse engineering experiment,dynamic data tracking and multi-stage comparative analysis are car-ried out for the news app TouTiao.It is found that there is a dynamic evolution from divergence to convergence in the topic of the algorithm recommendation.In the early stage,it aims to expand the user's interest boundary,and in the later stage,the recommended topics tend to be concentrated;in each stage,the posi-tive content occupies a dominant position in recommendation;with the increase of use frequency,the proportion of central media content and advertising content be-ing recommended tends to rise;besides,there are some structural differences in propensity of topic and emotional tendency recommended for different groups.Only when the algorithm power is moderately constrained and the user rights are effec-tively protected can the information dividend of artificial intelligence be maxi-mized under the premise of ethics.