Seminar Advances in Artificial Intelligence

Even though one of the main characteristics of ontologies has always been claimed to be their reusability, throughout this paper it will be shown that ontology reuse across a given domain is not a consolidated practice. This is confirmed through a statistical study on ontology reuse in the ontologies collected in Linked Open Vocabularies (LOV), in addition to a particular analysis of a use case. The results of the present work show that, when building an ontology, the heterogeneity between the needed conceptualisation and that of available ontologies, as well as the deficiencies in some of such ontologies (concerning documentation, licensing, etc.) are important obstacles for reusing ontologies of the same domain of the ontology under development. A possible approach to lessen these problems could be the creation of communities similar to open software ones in charge of developing and maintaining ontologies.

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3315457
Supervisor: Milene Santos Teixeira

The Resource Description Framework (RDF) is a standardized data model for representing machine processable knowledge within the Semantic Web. Increasingly large RDF datasets are published, e.g., driven by initiatives like the Linked Open Data movement. Traditional RDF representations are very verbose and not well suited for an efficient storage and exchange.
The article presents a binary RDF representation that addresses the issue. Based on a set of metrics that characterizes the skewed structure of real-world RDF data, an RDF representation format is developed that modularly partitions and efficiently represents three components of RDF datasets: Header information, a Dictionary, and the actual Triples structure (thus called HDT). An experimental evaluation shows that datasets in HDT format can be compacted by more than fifteen times as compared to current naive representations, improving both parsing and processing while keeping a consistent publication scheme. Specific compression techniques over HDT further improve these compression rates and prove to outperform existing compression solutions for efficient RDF exchange.

https://www.sciencedirect.com/science/article/pii/S1570826813000036
Supervisor: Birte Glimm

In order to deal with the dynamic nature of data streams, the latter is often segmented into so-called "windows" that represent static subparts of the streaming data. Typically, these windows are of fixed size and defined by a user independently of the actual streaming data. In this paper, a new type of data-dependent windows, called "frames", is introduced which actively consider and adapt to the processed data for the stream segmentation, thus providing more flexibility and expressiveness.

https://dl.acm.org/doi/pdf/10.1145/2933267.2933304
Supervisor: Moritz Illich

The Resource Description Framework (RDF) is a standardized data model for representing machine processable knowledge within the Semantic Web. Stream Reasoning deals with reasoning over streams of RDF data. This article presents a technique for incrementally maintaining logical consequences over windows of RDF data streams. The technique exploits time information to determine expired and new logical consequences. The provided experimental evidence shows that the approach significantly reduces the time required to compute valid inferences at each window change.

https://larkc.com/wp-content/uploads/2008/01/2010-Incremental-Reasoning-on-Streams-and-Rich-Background-Knowledge.pdf
Supervisor: Moritz Illich

The paper presents incremental evaluation algorithms to compute changes to materialized views in relational and deductive database systems, in response to changes (insertions, deletions, and updates) to the relations. The view definitions can be in SQL or Datalog, and may use UNION, negation, aggregation (e.g., SUM, MIN), linear recursion, and general recursion. The first algorithm is based on counting and tracks the number of alternative derivations (counts) for each derived tuple in a view. A second presented algorithm is called Delete and Rederive, DRed, and maintains views incrementally also for recursive views (negation and aggregation are permitted). 

https://dl.acm.org/doi/pdf/10.1145/170036.170066
Supervisor: Michael Welt

There is a large number of planning systems that work in different ways, which are based on search and are guided by heuristics. There are significantly more and more advanced techniques and heuristics for classic planning. In this seminar paper, therefore, the progression search and a technique are to be presented, such as how to use classic heuristics for search control in hierarchical planning.
There is a large number of different planning systems that are based on search guided by heuristics. There are significantly more and more advanced techniques and heuristics for classic planning. This seminar paper, presents progression search and a technique that allows for using heuristics from classical planning in hierarchical planning.

https://jair.org/index.php/jair/article/view/11282/26578
Supervisor: Conny Olz

For sophisticated reinforcement learning (RL) systems to interact usefully with real-world environments, we need to communicate complex goals to these systems. This work explores goals defined in terms of (non-expert) human preferences between pairs of trajectory segments. It is shown that this approach can effectively solve complex RL tasks without access to the reward function, including Atari games and simulated robot locomotion, while providing feedback on less than one percent of the agent's interactions with the environment. This reduces the cost of human oversight far enough that it can be practically applied to state-of-the-art RL systems. To demonstrate the flexibility of the approach, complex novel behaviors are successfully trained with about an hour of human time. These behaviors and environments are considerably more complex than any that have been previously learned from human feedback.

http://papers.neurips.cc/paper/7017-deep-reinforcement-learning-from-human-preferences.pdf
Supervisor: Jakob Karalus

The introduced framework abstracts Reinforcement Learning (RL) as a sequence modeling problem. This allows for drawing upon the simplicity and scalability of the Transformer architecture, and associated advances in language modeling such as GPT-x and BERT. In particular, the presented Decision Transformer is an architecture that casts the problem of RL as conditional sequence modeling. Unlike prior approaches to RL that fit value functions or compute policy gradients, Decision Transformer simply outputs the optimal actions by leveraging a causally masked Transformer. By conditioning an autoregressive model on the desired return (reward), past states, and actions, our Decision Transformer model can generate future actions that achieve the desired return. Despite its simplicity, Decision Transformer matches or exceeds the performance of state-of-the-art model-free offline RL baselines on Atari, OpenAI Gym, and Key-to-Door tasks.

https://arxiv.org/abs/2106.01345
Supervisor: Jakob Karalus

The aim of eXplainable Artificial Intelligence (XAI) is to design intelligent systems that can explain their predictions or recommendations to humans. Such systems are particularly desirable for therapeutic decision support, because physicians need to understand recommendations to have confidence in their application and to adapt them if required, e.g. in case of patient contraindication. The paper proposes an explainable and visual approach for decision support in antibiotic treatment, based on an ontology. The method consists of three steps: First, a tabular dataset is generated from the ontology, containing features defined on various domains and n-ary features. A preference model is then learned from patient profiles, antibiotic features and expert recommendations found in clinical practice guidelines. This model makes the implicit rationale of the expert explicit, including the way in which missing data was treated. As a third step, the preference model and its application to all antibiotics available on the market for a given clinical situation is visualized, using rainbow boxes, a recently developed technique for set visualization. The resulting preference model has an error rate of 3.5% on the learning data, and 5.2% on test data (10-fold validation). These findings suggest that the system can help physicians to prescribe antibiotics correctly, even for clinical situations not present in the guidelines (e.g. due to allergies or contraindications for the recommended treatment).

https://www.sciencedirect.com/science/article/pii/S1532046420300356?via%3Dihub
Supervisor: Milene Santos Teixeira

Explainability in Artificial Intelligence has been revived as a topic of active research by the need to demonstrate safety to users and gain their trust in the ‘how’ and ‘why’ of automated decisionmaking. Whilst a plethora of approaches have been developed for post-hoc explainability, only a few focus on how to use domain knowledge, and how it influences the understandability of global explanations from the users’ perspective. This paper shows how to use ontologies to create more understandable post-explanations of machine learning models. In particular, the approach builds on TREPAN, an algorithm that explains artificial neural networks by means of decision trees, which is extended to TREPAN Reloaded by including ontologies that model domain knowledge in the process of generating explanations. The results of a user study measure the understandability of decision trees through time and accuracy of responses as well as reported user confidence and understandability in relation to syntactic complexity of the trees. The user study considers domains where explanations are critical, namely finance and medicine. The results show that decision trees generated with our algorithm, taking into account domain knowledge, are more understandable than those generated by standard TREPAN without the use of ontologies.

https://ecai2020.eu/papers/666_paper.pdf
Supervisor: Felix Lindner

Understanding why a model makes a certain prediction can be as crucial as the prediction's accuracy in many applications. However, the highest accuracy for large modern datasets is often achieved by complex models that even experts struggle to interpret, such as ensemble or deep learning models, creating a tension between accuracy and interpretability. In response, various methods have recently been proposed to help users interpret the predictions of complex models, but it is often unclear how these methods are related and when one method is preferable over another. To address this problem, we present a unified framework for interpreting predictions, SHAP (SHapley Additive exPlanations). SHAP assigns each feature an importance value for a particular prediction. Its novel components include: (1) the identification of a new class of additive feature importance measures, and (2) theoretical results showing there is a unique solution in this class with a set of desirable properties. The new class unifies six existing methods, notable because several recent methods in the class lack the proposed desirable properties. Based on insights from this unification, we present new methods that show improved computational performance and/or better consistency with human intuition than previous approaches.

https://dl.acm.org/doi/pdf/10.5555/3295222.3295230
Supervisor: Amar Halilovic