Page 24 - Sirius_Annual_Report_2021

Team members
Basil Ell (program leader) Daniel Bakkelund
Egor V. Kostylev
Erik Bryhn Myklebust Ernesto Jimenez-Ruiz Evgeny Kharlamov
Gong Cheng
Ingrid Chieh Yu Martin Giese Jiaoyan Chen
Ole Magnus Holter Peyman Rasouli Roxana Pop Summaya Mumtaz
• In the context of classification, based on a use case that is relevant to the oil and gas industry, namely
that of excess inventory reduction, Daniel Bakkelund has developed theory and methodology for improved classification of interchangeable equipment, by inte- grating equipment structure awareness into classical methods for unsupervised machine learning [7]. Daniel will submit his PhD thesis in 2022.
• Jiaoyan Chen, Ernesto Jimenez-Ruiz, Ole Magnus Holter, Ian Horrocks and colleagues have developed an ontology embedding framework named OWL2Vec* that can embed symbolic knowledge in an OWL ontology into a vector space, so that the information can be consumed by machine learning algorithms. OWL2Vec* can be directly applied to ontology completion tasks such as subsumption prediction as well as to help address machine learning challenges, such as sample shortage, by injecting symbolic knowledge [8, 9].
• Actionable recourse (AR) techniques are a popular class of post-hoc interpretability approaches that help the users of ML models to obtain their desired decision from a machine learning model. Given an individual’s preferences, an AR recommends feasible changes to their corresponding input that lead to the desired outcome by the model. To generate realistic ARs, it is important to capture and exploit the domain’s information and the preferences of the users in the explanation process. Peyman Rasouli and Ingrid Chieh Yu are working on a model-agnostic framework that combines user/domain- level knowledge with model/data-level information to create plausible ARs that can guide individuals to obtain their desired decision from any ML classification and regression model in a simple and efficient manner.
• Current explainable artificial intelligence (XAI) techniques only rely on the observational data to analyze and explain the behavior of machine learning models.
To increase the comprehensibility and faithfulness
of explanations of ML models, hence, it is essential to exploit domain knowledge that bridges between the models and human concepts. Peyman Rasouli and Ingrid Chieh Yu aim to integrate domain knowledge (in the form of knowledge graphs and taxonomies) with structured/tabular data to provide more feasible, comprehensible, and faithful explanations.
• Gong Cheng, Evgeny Kharlamov investigated keyword-based exploration of knowledge graphs [10,11] and proposed a novel method to generate smart snippets or summaries of large-scale knowledge graphs. Baifan Zhou, Evgeny Kharlamov and colleagues from SIRIUS showed how to facilitate development
Within our research program, we have developed hybrid approaches, gained evidence for the benefits of hybrid approaches, and work towards developing novel hybrid approaches:
• Erik Bryhn Myklebust, Ernesto Jimenez-Ruiz, Jiaoyan Chen and colleagues have shown in the context of ecotoxicological effect prediction that the accuracy of predictions can be improved when domain knowledge is incorporated into the prediction model – see Figure 1 [2].
• Ole Magnus Holter and Basil Ell develop approaches that make use of domain knowledge in the context of semantic parsing of textual requirements. Their goal is to formally represent (parts of) the meaning of textual requirements, so that the meaning of requirements becomes more accessible to machines and the manage- ment of requirements can be improved [3].
• Egor V. Kostylev and colleagues study theoretical and practical connections between graph neural networks (GNNs), a modern structure-aware machine learning architecture, and classic logic-based knowledge representation formalisms. In particular, they designed a family of monotonic GNNs that allow for an efficient translation to Datalog logic-based language, and developed an efficient INDIGO system for knowledge graph completion [4, 5].
• In the context of a task relevant for the oil and gas industry, namely reservoir analogue identification, Summaya Mumtaz and Martin Giese have shown that a similarity measure based on the combination of domain knowledge (in the form of a taxonomy) with classical frequency-based features leads to significantly better results [6]. The disputation of her PhD thesis took place in November 2021.
24 | SIRIUS ANNUAL REPORT 2021