Contributions published at Dynamic and Distributed Information Systems (Abraham Bernstein)

The data produced by efforts such as life logging is commonly multi modal and can have manifold interrelations with itself as well as external information. Representing this data in such a way that these rich relations as well as all the different sources can be leveraged is a non-trivial undertaking. In this paper, we present the first iteration of LifeGraph, a Knowledge Graph for lifelogging data. LifeGraph aims at not only capturing all aspects of the data contained in a lifelog but also linking them to external, static knowledge bases in order to put the log as a whole as well as its individual entries into a broader context. In the Lifelog Search Challenge 2020, we show a first proof-of-concept implementation of LifeGraph as well as a retrieval system prototype which utilizes it to search the log for specific events.

Open Information Extraction is the process of extracting domain-independent triples from natural language text. This thesis assesses the performance of Open Information Extraction in real-life scenarios by comparing two state-of-the-art algorithms, namely Supervised-oie and Open IE 5.0, against each other and assesses how their reported performance diers from the one achieved on a real-life business corpus. The performance is measured with regards to precision, recall and runtime. The results suggest that there is a gap between the reported results and the ones achieved on the business corpus. Finally, an in-depth error assessment of the algorithms is conducted to suggest solutions to mitigate these errors and maximize both precision and recall.

In the information age we live today, it is absolutely crucial for organizations to provide an efficient means of storing, managing and retrieving their data. One approach to take on this challenge, which has been widely employed in practice, is faceted search. In corresponding search applications, data is structured using a faceted classification scheme, which offers great flexibility in comparison to traditional monohierarchical taxonomies. Social tagging, or social indexing comes into play if the classification process is a collaborative effort. In the context of the Digital Society Initative (DSI) - an academic platform at the University of Zurich - an application is required to retrieve members dependent on their disciplines of expertise. Such a Yellow Pages web application already exists, but has several undesired properties. This work re-implements this search application and tackles the issues of the former version by making use of a faceted classification scheme and by incorporating a social tagging mechanism. Further, a means of updating the Yellow Pages' data is implemented. In order to assess the usefulness and usability of the deployed application, a user survey is conducted. The re-implementation offers a strict superset of functionality compared to the predecessor version and users recognize the use cases and the flexibility offered by the faceted searching scheme and social tagging. Users generally think that navigating the new DSI Yellow Pages is straight forward.

We develop a taxonomy that categorizes HRI failure types and their impact on trust to structure the broad range of knowledge contributions. We further identify research gaps in order to support fellow researchers in the development of trustworthy robots. Studying trust repair in HRI has only recently been given more interest and we propose a taxonomy of potential trust violations and suitable repair strategies to support researchers during the development of interaction scenarios. The taxonomy distinguishes four failure types: Design, System, Expectation, and User failures and outlines potential mitigation strategies. Based on these failures, strategies for autonomous failure detection and repair are presented, employing explanation, verification and validation techniques. Finally, a research agenda for HRI is outlined, discussing identified gaps related to the relation of failures and HR-trust.

null

null

null

null

The natural language expression “near” describes spatial proximity. However, the interpretation of this expression depends on the context. In this thesis, we investigate how a context-dependent model for “near” can be formulated. For doing so, we investigate the following questions: (i) what is the relevant contextual information for “near”? (ii) how does the identified information influence the interpretation of near? To answer these questions, the research conducted consists in identifying the relevant contextual information from the literature. Subsequently, different contextualized nearness models are formulated to evaluate the influence of the context on “near”. To train the contextualized nearness models, the necessary data is extracted from the geograph.co.uk corpus. The data is extracted using a probabilistic semantic geo-parser, which we built on the basis of the insights gained in this thesis.

Background: Stress is a complex phenomenon that may have a negative influence on health and well-being; consequently, it plays a pivotal role in mental health. Although the incidence of mental disorders has been continuously rising, development of prevention and treatment methods has been rather slow. Through the ubiquitous presence of smartphones and wearable devices, people can monitor stress parameters in everyday life. However, the reliability and validity of such monitoring are still unsatisfactory.Methods: The aim of this trial is to find a relationship between psychological stress and saliva cortisol levels on the one hand and physiological parameters measured by smartphones in combination with a commercially available wearable device on the other. Participants include cohorts of individuals with and without a psychiatric disorder. The study is conducted in two settings: one naturalistic and one a controlled laboratory environment, combining ecological momentary assessment (EMA) and digital phenotyping (DP). EMA is used for the assessment of challenging and stressful situations coincidentally happening during a whole observation week. DP is used during a controlled stress situation with the Trier Social Stress Test (TSST) as a standardized psychobiological paradigm. Initially, participants undergo a complete psychological screening and profiling using a standardized psychometric test battery. EMA uses a smartphone application, and the participants keep a diary about their daily routine, activities, well-being, sleep, and difficult and stressful situations they may encounter. DP is conducted through wearable devices able to continuously monitor physiological parameters (i.e., heart rate, heart rate variability, skin conductivity, temperature, movement and acceleration). Additionally, saliva cortisol samples are repeatedly taken. The TSST is conducted with continuous measurement of the same parameters measured during the EMA.Discussion: We aim to identify valid and reliable digital biomarkers for stress and stress reactions. Furthermore, we expect to find a way of early detection of psychological stress in order to evolve new opportunities for interventions reducing stress. That may allow us to find new ways of treating and preventing mental disorders.Trial Registration: The competing ethics committee of the Canton of Zurich, Switzerland, approved the study protocol V05.1 May 28, 2019 BASEC: 2019-00814; the trial was registered at ClinicalTrials.gov NCT04100213 on September 19, 2019.

Access to diverse sources of news and information is more important than ever in this time of global crisis, yet far too often, people can Bnd themselves diving into ‘rabbit holes’ of information and opinion that are hard to escape. Here, the following authors provide an indepth assessment of how algorithmic recommendation systems for news can play a more constructive role in a diverse media landscape.

Free-ranging wild boar (Sus scrofa) in Switzerland: casual observations and model-based projections during open and closed season for hunting Wild boar (i.e., Sus scrofa) are susceptible to a range of diseases that can be transmitted to domestic pigs. Assessing the potential risk of transmission-related events involves identifying where wild boar occur in Switzerland and where they still can colonize. It also involves identifying zones where piggeries are dense. In the work presented here, the distribution of wild boar in Switzerland was projected from grid data as probabilities of presence using an approach based on statistical modeling, separately for closed and open season for hunting. The predicted probabilities of wild boar presence were related to the density of piggeries in the six agricultural zones. The resulting maps show how the potential risk of transmission-related events, as a proxy for disease transmission, is distributed in Switzerland. Wild boar presence data consisted of hunting data and casual observations recorded from September 2011 to February 2018 at the coordinate level. They were obtained from all 16 Swiss cantons maintaining a license hunting system plus Solothurn (for 2017) and Zurich, as well as from info fauna. The probability of wild boar occurrence was high (> 0.7) in Jura, the valleys of the Southern Alps, the Rhone Valley down the river from Martigny, and the Rhine Valley down the river from Bündner Herrschaft; it was fair (0.5–0.7) in the Swiss Plateau. These regions broadly overlap agricultural zones with a high density of piggeries. Patches of perennially suitable, but currently not colonized habitat were found in the cantons of Berne, Obwalden, Uri, Schwyz, Glarus, and Grisons. The probability of wild boar occurrence across the entire study area, including the Alps, increased by 12% during closed season for hunting. The results were discussed with reference to similar studies.

Nearness expressions such as "near"are context-dependent spatial relations and are subject to the context variability effect. Depending on the provided context, "near"has a different semantic extension. We perform a literature review to identify the effect of context on "near". To integrate the insights from different disciplines, we apply Turney's contextualization framework which distinguishes between two types of features: primary and contextual. Primary features are the qualitative and quantitative distance measures and contextual features are the context factors used to determine a threshold on the nearness measurements. Additionally, we identify the appropriate features for different spatial tasks discussed in the literature. By doing so, we seek to build a foundation for a context-dependent model for "near".

As people increasingly rely on online media and recommender systems to consume information, engage in debates and form their political opinions, the design goals of online media and news recommenders have wide implications for the political and social processes that take place online and offline. Current recommender systems have been observed to promote personalization and more effective forms of informing, but also to narrow the user’s exposure to diverse content. Concerns about echo-chambers and filter bubbles highlight the importance of design metrics that can successfully strike a balance between accurate recommendations that respond to individual information needs and preferences, while at the same time addressing concerns about missing out important information, context and the broader cultural and political diversity in the news, as well as fairness. A broader, more sophisticated vision of the future of personalized recommenders needs to be formed - a vision that can only be developed as the result of a collaborative effort by different areas of academic research (media studies, computer science, law and legal philosophy, communication science, political philosophy, and democratic theory). The proposed workshop will set first steps to develop such a much needed vision on the role of recommender systems on the democratic role of the media and define the guidelines as well as a manifesto for future research and long-term goals for the emerging topic of fairness, diversity, and personalization in recommender systems.

Stream Processing (SP) is an important Big Data technology enabling continuous querying of data streams. The stream setting offers the opportunity to exploit synergies and, theoretically, share the access and processing costs between multiple different collaborators. But what should be the monetary contribution of each consumer when they do not trust each other and have varying valuations of the differing outcomes? In this article, we present Collaborative Stream Processing (CSP), a model where the costs, which are set exogenously by providers, are shared between multiple consumers, the collaborators. For this, we identify three important requirements for CSP to establish trust between the collaborators and propose a CSP al- gorithm, ENCSPA, adhering to these requirements. Based on the collaborators’ outcome valuations and the costs of the raw data streams, ENCSPA computes the payment for each collaborator. At the same time, ENCSPA ensures that no collaborator has an incentive to manipulate the system by providing misinformation about her/his value, budget, or time limit. We show that ENCSPA can calculate payments in a reasonable amount of time for up to one thousand collaborators.

Epidemic spreading occurs among animals, humans, or computers and causes substantial societal, personal, or economic losses if left undetected. Based on known temporal contact networks, we propose an outbreak detection method that identifies a small set of nodes such that the likelihood of detecting recent outbreaks is maximal. The two-step procedure involves i) simulating spreading scenarios from all possible seed configurations and ii) greedily selecting nodes for monitoring in order to maximize the detection likelihood. We find that the detection likelihood is a submodular set function for which it has been proven that greedy optimization attains at least 63% of the optimal (intractable) solution. The results show that the proposed method detects more outbreaks than benchmark methods suggested recently and is robust against badly chosen parameters. In addition, our method can be used for out- break source detection. A limitation of this method is its heavy use of computational resources. However, for large graphs the method could be easily parallelized.

Studying the evolution of knowledge graphs has become an important topic of current research. For that purpose, this thesis provides a foundation by contributing a framework that creates historic snapshots of the Wikidata knowledge graph. In order to save resources, the framework extracts a sample out of the original graph and generates the snapshots based on that sample. The behavior and biases of different traversal-based sampling techniques are analyzed and they agree with previous observations by related work on sampling. This work further describes the types of revisions and how they can be undone in order to create a sequence of versions of the sampled graph in earlier stages of its history. Finally, it demonstrates how the snapshots are returned in a standard RDF format.