Defusing discrimination traps of AI's
Numerous studies have shown that suggestions created by recommender systems (RSs) can discriminate against certain demographic groups (e.g., according to gender or race). The TIMELY project targets this problem in the context of gender bias in job RSs, following an interdisciplinary approach that amalgamates expertise from machine learning and nondiscrimination law.
Assoc.-Prof.in Dr.in Elisabeth Greif
Assist.-Prof. Dr. Navid Rekabsaz
The core technological component of TIMELY is a novel solution which we refer to as con- tentdebiased RS (CDRS). The proposed job CDRS includes an approach to identify po- tentially discriminatory wordings in job de- scriptions, using deep learning-based natural language processing (NLP) techniques. The resulting debiasing embeddings can be inte- grated into most content-based RSs, making the approach adaptable to other text-based recommendation domains.
Both the NLP-based identification of potentially discriminatory wordings and the CDRS are strongly intertwined with the addressed legal research aspects: mapping the legal framework for job advertising according to EU non-discrimination law will enable us to assess which of the wordings identified by the NLP component are indeed discriminatory and use this information to improve the NLP and CDRS components.
We will validate the devised CDRS solution in terms of accuracy, gender fairness, and user experi- ence, and additionally study how the results can be used to support litigation for discrimination in court. In addition to the machine learning and law expertise of the PI and co-PIs, TIMELY is also supported by Jobiqo GmbH, a provider of job platforms that offer search and recommendation functionality. Jobiqo will provide both financial support and access to data of their customers.
Prof. Schedl is leading the group Human-centered AI at the AI Lab of the LIT as well as the group Multimedia Mining and Search (MMS) at the Institute of Computational Perception at the JKU. His areas of expertise include recommender systems, user mod- eling, information retrieval, machine learning, natural language processing, multimedia, data analysis, and web mining.
He has a passion for teaching and regularly gives courses at the JKU as well as on guest lecturing stays, among others at Uni- versitat Pompeu Fabra Barcelona, Queen Mary University of London, and Kungliga Tekniska Högskolan Stockholm. He also serves as consultant on the topics mentioned above.
In his various projects he is also collaborating with industry, for instance with Siemens, Spotify, and Deezer. His research is as well interessting for the art world as he as recently shown at the Ars Electronica Festival 2021, opens an external URL in a new window.