ORGANISATION/COMPANYUniversitat Pompeu Fabra - ETIC
RESEARCH FIELDComputer science
RESEARCHER PROFILEFirst Stage Researcher (R1)
APPLICATION DEADLINE02/03/2021 03:00 - Europe/Athens
LOCATIONSpain › Barcelona
TYPE OF CONTRACTTemporary
HOURS PER WEEK20
OFFER STARTING DATE01/05/2021
This is a short-term research assistant position on the topic of algorithmic fairness and debiasing of graph embeddings. Specifically, the work consists on performing research along two axes -- first, to identify ways in which the use of graph embeddings in tasks such as node classification, link prediction, and graph-based recommendations, may introduce biases due to correlations between nodes’ personal, sensitive attributes and their connections to other nodes in the graph, and two, to develop new methods that can mitigate such biases when detected without severely impacting the performance of embedding algorithms along traditional measures such as accuracy.
Graph embeddings are continuous, low-dimensional, vector representations of nodes estimated from the graph topology as well as nodes’ personal attributes/metadata. Graph embeddings have increasingly been used in downstream tasks such as node classifications (eg. predicting the political affiliations of users of a social network), link prediction (eg. predicting the presence/absence of an edge between a given pair of nodes), and recommendation systems. This is because metadata can enhance graph learning models and graph topology can be used for regularization in supervised learning.
Methods to generate these embeddings typically rely on the homophily assumption. That is, they assume that nodes connected by a link are more likely to have similar attribute values and belong to the same class compared to nodes that are not connected by a link (Peel et al., Science Advances 2017). This may potentially introduce unwanted biases in some cases, for instance when fraudsters or inauthentic accounts in social media "camouflage" themselves by following many authentic accounts.
If users are segregated into groups and offered or excluded different products, services, or prices on the basis of their sensitive attributes (eg. gender, location), or their connections to other users of a protected group, this can potentially lead to indirect discrimination by association. Moreover, by taking a universalizing lens (eg. graph convolution networks that create node representations by aggregating attribute values from its neighbours), embeddings algorithmically redefine the relationship between the individual and the collective, blending preference, similarity, and identity (Latour et al., British Journal of Sociology 2012).
We want to obfuscate the embeddings during construction such that when they are used in a black-box fashion in a downstream task, information about sensitive attributes does not leak through. This is not as straightforward as simply choosing to not use sensitive attributes during training (disparate treatment) because this may be indirectly captured by the connections between nodes. Mitigation procedures must first analyze the potentially causal influences between attributes and graph topology and then provably debias topology embeddings from attribute embeddings to enable fairness at the subgroup and individual levels.
The goal of the research is to determine different ways and the extent to which bias may occur in graph embeddings by conducting experiments on synthetic graphs (to highlight limitations of existing approaches) as well as real-world datasets (to examine their impacts), and then design alternative methods for mitigating these potential harms.
This includes at least the following tasks: literature review; conceptualization and development of algorithms; creation of reference implementations; data collection and curation; experimentation; writing and presentation of results.
This job will be conducted at the Web Science and Social Computing Group.
Required Research Experiences
RESEARCH FIELDComputer science
YEARS OF RESEARCH EXPERIENCE1 - 4
REQUIRED EDUCATION LEVELComputer science: Master Degree or equivalent
REQUIRED LANGUAGESENGLISH: Excellent
- Two years of research experience:
Computer Science - Information Technology.
EURAXESS offer ID: 604551
The responsibility for the jobs published on this website, including the job description, lies entirely with the publishing institutions. The application is handled uniquely by the employer, who is also fully responsible for the recruitment and selection processes.
Please contact email@example.com if you wish to download all jobs in XML.