Fairness in AI-powered Search
Algorithmic bias and unfairness has become a large issue as AI and ML are being rapidly used in real-world.
What is “Unfair”?
In the context of decision-making, fairness is
- the absence of any prejudice or favouritism toward a subject, regardless of their intrinsic or acquired attributes.
- It can be unfair if an algorithm is much more likely to make a mistake about you than about others
Why might machine learning be unfair?
The world is unfair
- And a trained model simply reflects this
the power law suggests that a small number of head queries account for a disproportionately high amount of total traffic, while a large number of tail queries each contribute a tiny amount of traffic.
For many of those queries, engagement tends to be concentrated in a handful of top-ranked results, meaning they tend to be highly optimized already. As a result, the size of training data can overstate its utility. Your ML model, in its inherent nature, is designed to optimize for the most common patterns in the training data. This leads to the model being skewed towards the dominant traffic, typically the head queries, given their high frequency and the concentrated…
