Award for thesis on biased book recommendation systems

Julia Sudnik won the Amsterdam Data Science thesis award. She did an internship at CWI on ‘The effect of feedback loops on fairness in rankings from recommendation systems’.

Publication date
3 Jan 2023

Recommender systems are information filtering systems that provide suggestions for items that are most pertinent to a particular user. They refer to various decision-making processes, such as what product to purchase, what music to listen to, or what online news to read. Quite handy when you need to choose an item from a potentially overwhelming number of items that are offered.

But recommender systems can be biased. Bias in recommender systems may negatively affect users as well as producers. In recent years, many examples have been reported of recommender systems that discriminate against, for example, people of certain genders or race. More knowledge is needed about how to measure and mitigate bias in recommender systems.

Making bias visible

Recently, there has been substantial research undertaken on the role of fairness in recommendation systems. Sudniks thesis draws attention to variables affecting fairness evaluation of a recommendation system such as chosen definition of fairness, learning algorithm in use, ranking positions considered, and the amount of feedback loops already computed. With a focus on book recommendations and female writers as the minority groups, Sudnik has designed a clever experimental set-up with several models and fairness metrics.

The reviewers from Amsterdam Data Science were impressed by the work and considered the choice of methods to be excellent. “The thesis takes a clear approach to make such a bias visible, and provides an insightful discussion of the implications of this work by the use of different definitions of fairness, and providing detailed insights on the tradeoffs and factors to consider to address such complex issues.”

Julia Sudnik received the Amsterdam Data Science thesis award.
Julia Sudnik (second from the right) received the Amsterdam Data Science thesis award.

Current datasets are insufficient

The research took place in the cultural heritage sector (libraries). This sector is hesitant in its use of Artificial Intelligence for fear of introducing bias in their services, thus undermining their societal role as a trusted source of information. Sudniks study helps them in choosing algorithms that fit best with their goals (rather than ‘just’ the best performing algorithm).

Her thesis has pointed out that research datasets that are currently widely-used to study bias in recommender systems, are insufficient: they don’t contain the necessary sensitive information. Sudnik has shown how to extend such a dataset in a reliable and transparent manner, thus making it suitable for various types of bias detection. Secondly, she has demonstrated that data selection methods that are widely used in the field, are biased themselves – in particular the common practice of reducing sparseness in research datasets.
Finally, Sudnik showed that different fairness metrics lead to different conclusions regarding
fairness of algorithms. While this was known theoretically, very few studies have shown this