Modern algorithms, which, for example, make further suggestions for similar content based on the videos viewed, can predict the gender of the user with an accuracy of up to 72 percent. Other demographic data can also be determined with great accuracy.
Questionable in terms of data protection – which is why researchers at the Institute for Computational Perception are now looking at alternative methods of how online platforms can suggest attractive content to users without violating their privacy. “This is not a program that users can switch to, but the offer is aimed at companies,” says Markus Schedl, who is part of the research team.
Specifically, the JKU scientists developed a method in which two networks work together. One of them is engaged in the task of finding music or movies that the user might like. The second network, which adjusts the parameters step by step, is new. The recommendations remain similar, but it is difficult to draw conclusions about the users.
The prediction accuracy for the socio-demographic data of the users was reduced from the initially mentioned 72 percent for music suggestions to 57 percent and for films to 62 percent. When checking the results, the JKU researchers also discovered a positive side effect: With the new method, the algorithms make fewer stereotypical suggestions – such as romance films for women and action films for men. “Thus, our approach also counteracts the formation of filter bubbles on the Internet,” says Schedl.
The knowledge gained and the resulting algorithm are available free of charge. Private companies can therefore freely experiment with it.
Source: Nachrichten
