Refine
Document Type
- Doctoral Thesis (3)
- Master's Thesis (1)
Social networks are ubiquitous structures that we generate and enrich every-day while connecting with people through social media platforms, emails, and any other type of interaction. While these structures are intangible to us, they carry important information. For instance, the political leaning of our friends can be a proxy to identify our own political preferences. Similarly, the credit score of our friends can be decisive in the approval or rejection of our own loans. This explanatory power is being leveraged in public policy, business decision-making and scientific research because it helps machine learning techniques to make accurate predictions. However, these generalizations often benefit the majority of people who shape the general structure of the network, and put in disadvantage under-represented groups by limiting their resources and opportunities. Therefore it is crucial to first understand how social networks form to then verify to what extent their mechanisms of edge formation contribute to reinforce social inequalities in machine learning algorithms.
To this end, in the first part of this thesis, I propose HopRank and Janus two methods to characterize the mechanisms of edge formation in real-world undirected social networks. HopRank is a model of information foraging on networks. Its key component is a biased random walker based on transition probabilities between k-hop neighborhoods. Janus is a Bayesian framework that allows to identify and rank plausible hypotheses of edge formation in cases where nodes possess additional information. In the second part of this thesis, I investigate the implications of these mechanisms - that explain edge formation in social networks - on machine learning. Specifically, I study the influence of homophily, preferential attachment, edge density, fraction of inorities, and the directionality of links on both performance and bias of collective classification, and on the visibility of minorities in top-k ranks. My findings demonstrate a strong correlation between network structure and machine learning outcomes. This suggests that systematic discrimination against certain people can be: (i) anticipated by the type of network, and (ii) mitigated by connecting strategically in the network.
Semantic Web technologies have been recognized to be key for the integration of distributed and heterogeneous data sources on the Web, as they provide means to define typed links between resources in a dynamic manner and following the principles of dataspaces. The widespread adoption of these technologies in the last years led to a large volume and variety of data sets published as machine-readable RDF data, that once linked constitute the so-called Web of Data. Given the large scale of the data, these links are typically generated by computational methods that given a set of RDF data sets, analyze their content and identify the entities and schema elements that should be connected via the links. Analogously to any other kind of data, in order to be truly useful and ready to be consumed, links need to comply with the criteria of high quality data (e.g., syntactically and semantically accurate, consistent, up-to-date). Despite the progress in the field of machine learning, human intelligence is still essential in the quest for high quality links: humans can train algorithms by labeling reference examples, validate the output of algorithms to verify their performance on a data set basis, as well as augment the resulting set of links. Humans —especially expert humans, however, have limited availability. Hence, extending data quality management processes from data owners/publishers to a broader audience can significantly improve the data quality management life cycle.
Recent advances in human computation and peer-production technologies opened new avenues for human-machine data management techniques, allowing to involve non-experts in certain tasks and providing methods for cooperative approaches. The research work presented in this thesis takes advantage of such technologies and investigates human-machine methods that aim at facilitating link quality management in the Semantic Web. Firstly, and focusing on the dimension of link accuracy, a method for crowdsourcing ontology alignment is presented. This method, also applicable to entities, is implemented as a complement to automatic ontology alignment algorithms. Secondly, novel measures for the dimension of information gain facilitated by the links are introduced. These entropy-centric measures provide data managers with information about the extent the entities in the linked data set gain information in terms of entity description, connectivity and schema heterogeneity. Thirdly, taking Wikidata —the most successful case of a linked data set curated, linked and maintained by a community of humans and bots— as a case study, we apply descriptive and predictive data mining techniques to study participation inequality and user attrition. Our findings and method can help community managers make decisions on when/how to intervene with user retention plans. Lastly, an ontology to model the history of crowd contributions across marketplaces is presented. While the field of human-machine data management poses complex social and technical challenges, the work in this thesis aims to contribute to the development of this still emerging field.
Within the field of Business Process Management, business rules are commonly used to model company decision logic and govern allowed company behavior. An exemplary business rule in the financial sector could be for example:
”A customer with a mental condition is not creditworthy”. Business rules are
usually created and maintained collaboratively and over time. In this setting,
modelling errors can occur frequently. A challenging problem in this context is
that of inconsistency, i.e., contradictory rules which cannot hold at the same
time. For instance, regarding the exemplary rule above, an inconsistency would
arise if a (second) modeller entered an additional rule: ”A customer with a mental condition is always creditworthy”, as the two rules cannot hold at the same
time. In this thesis, we investigate how to handle such inconsistencies in business
rule bases. In particular, we develop methods and techniques for the detection,
analysis and resolution of inconsistencies in business rule bases
This thesis focuses on approximate inference in assumption-based argumentation frameworks. Argumentation provides a significant idea in the computerization of theoretical and practical reasoning in AI. And it has a close connection with AI, engaging in arguments to perform scientific reasoning. The fundamental approach in this field is abstract argumentation frameworks developed by Dung. Assumption-based argumentation can be regarded as an instance of abstract argumentation with structured arguments. When facing a large scale of data, a challenge of reasoning in assumption-based argumentation is how to construct arguments and resolve attacks over a given claim with minimal cost of computation and acceptable accuracy at the same time. This thesis proposes and investigates approximate methods that randomly select and construct samples of frameworks based on graphical dispute derivations to solve this problem. The presented approach aims to improve reasoning performance and get an acceptable trade-off between computational time and accuracy. The evaluation shows that for reasoning in assumption-based argumentation, in general, the running time is reduced with the cost of slightly low accuracy by randomly sampling and constructing inference rules for potential arguments over a query.
