Auditing Autocomplete: Suggestion Networks and Recursive Algorithm Interrogation
Auditing Autocomplete: Suggestion Networks and Recursive Algorithm Interrogation
复制标题
审核自动完成：建议网络和递归算法询问
DOI：
10.1145/3292522.3326047
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
Christo Wilson
中科院分区：
文献类型：
--
作者：
Ronald E. Robertson;Shan Jiang;D. Lazer;Christo Wilson
自动完成算法，通过设计，引导查询。当用户提供根输入（例如搜索查询）时，这些算法动态地检索、策划和呈现相关输入（例如搜索建议）的列表。虽然在线平台无处不在，但由于缺乏研究来解决其产出的短暂性及其运作的不透明性，人们对调查的透明度和问责制表示关切。在这里，我们介绍递归算法询问（RAI），一个广度优先的搜索方法，通过递归提交一个根查询及其子建议，以创建一个网络的算法关联审计自动完成。在2018年夏天，我们使用RAI对Google和Bing上的自动完成进行了纵向审计，使用了一组集中的根查询--38位竞选连任的美国州长的名字。在搜索引擎之间进行比较，我们发现两个搜索引擎上较长和较低排名的建议的周转率更高，Google建议中社交媒体网站的流行率更高，Bing建议中被归类为脏话或负面情绪的单词的流行率更高，以及跨越我们大多数根查询的周期性冲击。我们开源我们的代码进行RAI，并讨论如何将其应用到其他平台，主题和设置。
Autocomplete algorithms, by design, steer inquiry. When a user provides a root input, such as a search query, these algorithms dynamically retrieve, curate, and present a list of related inputs, such as search suggestions. Although ubiquitous in online platforms, a lack of research addressing the ephemerality of their outputs and the opacity of their functioning raises concerns of transparency and accountability on where inquiry is steered. Here, we introduce recursive algorithm interrogation (RAI), a breadth-first search method for auditing autocomplete by recursively submitting a root query and its child suggestions to create a network of algorithmic associations. We used RAI to conduct a longitudinal audit of autocomplete on Google and Bing using a focused set of root queries -- the names of 38 US governors who were up for reelection -- during the summer of 2018. Comparing across search engines, we found a higher turnover rate among longer and lower ranked suggestions on both search engines, a higher prevalence of social media websites in Google's suggestions, a higher prevalence of words classified as a swear or a negative emotion in Bing's suggestions, and periodic shocks that spanned across most of our root queries. We open source our code for conducting RAI and discuss how it could be applied to other platforms, topics, and settings.