Cornell University computer scientists have developed a tool which can improve the fairness of online search rankings, without sacrificing relevance or usefulness.
For much of the world, Google Search is the entrance to the rest of the internet (other search engines are available). With most users clicking the results towards the top of the first page of returned search rankings, rarely venturing beyond that first page, the vast majority of options remain hidden, creating a potential for bias in everything from e-commerce to recruitment.
In an award-winning new paper, computer scientists from Cornell University describe a new tool (‘FairCo’) which they hope could improve the fairness of online rankings without having an adverse effect on relevance or usefulness.
“If you could examine all your choices equally and then decide what to pick, that may be considered ideal. But since we can’t do that, rankings become a crucial interface to navigate these choices,” said co-first author of the study, PhD student Ashudeep Singh.
“For example, many YouTubers will post videos of the same recipe, but some of them get seen way more than others, even though they might be very similar. This happens because of the way search results are presented to us. We generally go down the ranking linearly and our attention drops off fast.”
As existing algorithms seek the most relevant items to searchers, and searchers tend to choose from the very top ranked items in a list, very small differences in relevance can lead to massive discrepancies in exposure. However, the new method proposed by the researchers gives approximately equal exposure to equally relevant choices and avoids preferential treatment for items which are already ranked highly.
This adaptation corrects the unfairness which is inherent in existing algorithms, which have been criticised for exacerbating inequality and political polarisation while limiting individual choice and making it difficult for small websites to compete with bigger, more established websites.
“What ranking systems do is they allocate exposure. How do we make sure that everybody receives their fair share of exposure?” said senior author Professor Thorsten Joachims. “What constitutes fairness is probably very different in, say, an e-commerce system and a system that ranks résumés for a job opening. We came up with computational tools that let you specify fairness criteria, as well as the algorithm that will probably enforce them.”
The algorithm takes the form of a controller which integrates unbiased estimators for both fairness and utility, adapting as more data are made available.
For instance, if applied to an e-commerce system, the tool could be used to ensure that if a product is liked by 30 per cent of searchers, its exposure is based more proportionally on that level of interest. If applied to a database of CVs, safeguards could be applied to ensure that it does not discriminate by protected characteristics.
Evaluations based on a semi-synthetic news dataset and real-world film preference data showed that the tool is effective at eliminating bias, while still ranking useful results highly.