The Challenges of Developing Successful Algorithms

The final speaker in this AoIR 2015 session is Anja Bechmann, whose focus is on algorithms from the designer's point of view. Often, users are portrayed as the victims of victorious algorithms – but algorithms are only powerful if they have the right data to process.

We assume that algorithms are streamlining and simplifying activities and are sensitive to our wants and needs, and to do so most effectively they need to interact and interoperate with each other; algorithmic identity thus rests not on an essentialist but on a constructivist approach as identity is enacted when databases meet databases through algorithms.

Anja demonstrates this through the example of the Danish algorithm project Det Sociale Bibliotek (The Social Library), which draws on Facebook users' newsfeeds to present them with relevant library materials. It does not draw simply on their own posts, as the newsfeed also represents the interests of users' friends, and thus of their social environment.

The algorithm draws on this content and identifies the ten most important non-trivial terms; these are weighted through a comparison with the user's own publicly stated interests, and the algorithm then makes a range of suggestions for interesting library materials. Users can then like or dislike such suggestions, and these responses also flow into the further personalisation of these recommendations.

Interviews show that such recommendations themselves are not specific enough: the algorithm requires further improvement. The relational status of connections with specific Facebook friends needs to be further incorporated – we may not want recommendations based on the interests of all of our friends, for example; and recommendations should not necessarily focus on what is already in the newsfeed, but suggest other ideas that are similar to what's already there.

Algorithmic results are thus like Schrödinger's cat, whose status can only be determined by opening the box: the quality of an algorithm can only be determined by asking users about their experiences in using them.