Refine
Document Type
- Doctoral Thesis (11) (remove)
Keywords
- Maschinelles Lernen (2)
- Annotation (1)
- Augenbewegung (1)
- Auslese (1)
- Auswahl (1)
- Bipartiter Graph (1)
- Blickbewegung (1)
- Enhanced Representation (1)
- Eye Tracking (1)
- Eyetracking (1)
Institute
- Institute for Web Science and Technologies (11) (remove)
In this thesis, I study the spectral characteristics of large dynamic networks and formulate the spectral evolution model. The spectral evolution model applies to networks that evolve over time, and describes their spectral decompositions such as the eigenvalue and singular value decomposition. The spectral evolution model states that over time, the eigenvalues of a network change while its eigenvectors stay approximately constant.
I validate the spectral evolution model empirically on over a hundred network datasets, and theoretically by showing that it generalizes arncertain number of known link prediction functions, including graph kernels, path counting methods, rank reduction and triangle closing. The collection of datasets I use contains 118 distinct network datasets. One dataset, the signed social network of the Slashdot Zoo, was specifically extracted during work on this thesis. I also show that the spectral evolution model can be understood as a generalization of the preferential attachment model, if we consider growth in latent dimensions of a network individually. As applications of the spectral evolution model, I introduce two new link prediction algorithms that can be used for recommender systems, search engines, collaborative filtering, rating prediction, link sign prediction and more.
The first link prediction algorithm reduces to a one-dimensional curve fitting problem from which a spectral transformation is learned. The second method uses extrapolation of eigenvalues to predict future eigenvalues. As special cases, I show that the spectral evolution model applies to directed, undirected, weighted, unweighted, signed and bipartite networks. For signed graphs, I introduce new applications of the Laplacian matrix for graph drawing, spectral clustering, and describe new Laplacian graph kernels. I also define the algebraic conflict, a measure of the conflict present in a signed graph based on the signed graph Laplacian. I describe the problem of link sign prediction spectrally, and introduce the signed resistance distance. For bipartite and directed graphs, I introduce the hyperbolic sine and odd Neumann kernels, which generalize the exponential and Neumann kernels for undirected unipartite graphs. I show that the problem of directed and bipartite link prediction are related by the fact that both can be solved by considering spectral evolution in the singular value decomposition.
Tagging systems are intriguing dynamic systems, in which users collaboratively index resources with the so-called tags. In order to leverage the full potential of tagging systems, it is important to understand the relationship between the micro-level behavior of the individual users and the macro-level properties of the whole tagging system. In this thesis, we present the Epistemic Dynamic Model, which tries to bridge this gap between the micro-level behavior and the macro-level properties by developing a theory of tagging systems. The model is based on the assumption that the combined influence of the shared background knowledge of the users and the imitation of tag recommendations are sufficient for explaining the emergence of the tag frequency distribution and the vocabulary growth in tagging systems. Both macro-level properties of tagging systems are closely related to the emergence of the shared community vocabulary. rnrnWith the help of the Epistemic Dynamic Model, we show that the general shape of the tag frequency distribution and of the vocabulary growth have their origin in the shared background knowledge of the users. Tag recommendations can then be used for selectively influencing this general shape. In this thesis, we especially concentrate on studying the influence of recommending a set of popular tags. Recommending popular tags adds a feedback mechanism between the vocabularies of individual users that increases the inter-indexer consistency of the tag assignments. How does this influence the indexing quality in a tagging system? For this purpose, we investigate a methodology for measuring the inter-resource consistency of tag assignments. The inter-resource consistency is an indicator of the indexing quality, which positively correlates with the precision and recall of query results. It measures the degree to which the tag vectors of indexed resources reflect how the users perceive the similarity between resources. We argue with our model, and show it with a user experiment, that recommending popular tags decreases the inter-resource consistency in a tagging system. Furthermore, we show that recommending the user his/her previously used tags helps to increase the inter-resource consistency. Our measure of the inter-resource consistency complements existing measures for the evaluation and comparison of tag recommendation algorithms, moving the focus to evaluating their influence on the indexing quality.
The availability of digital cameras and the possibility to take photos at no cost lead to an increasing amount of digital photos online and on private computers. The pure amount of data makes approaches that support users in the administration of the photo necessary. As the automatic understanding of photo content is still an unsolved task, metadata is needed for supporting administrative tasks like search or photo work such as the generation of photo books. Meta-information textually describes the depicted scene or consists of information on how good or interesting a photo is.
In this thesis, an approach for creating meta-information without additional effort for the user is investigated. Eye tracking data is used to measure the human visual attention. This attention is analyzed with the objective of information creation in the form of metadata. The gaze paths of users working with photos are recorded, for example, while they are searching for photos or while they are just viewing photo collections.
Eye tracking hardware is developing fast within the last years. Because of falling prices for sensor hardware such as cameras and more competition on the eye tracker market, the prices are falling, and the usability is increasing. It can be assumed that eye tracking technology can soon be used in everyday devices such as laptops or mobile phones. The exploitation of data, recorded in the background while the user is performing daily tasks with photos, has great potential to generate information without additional effort for the users.
The first part of this work deals with the labeling of image region by means of gaze data for describing the depicted scenes in detail. Labeling takes place by assigning object names to specific photo regions. In total, three experiments were conducted for investigating the quality of these assignments in different contexts. In the first experiment, users decided whether a given object can be seen on a photo by pressing a button. In the second study, participants searched for specific photos in an image search application. In the third experiment, gaze data was collected from users playing a game with the task to classify photos regarding given categories. The results of the experiments showed that gaze-based region labeling outperforms baseline approaches in various contexts. In the second part, most important photos in a collection of photos are identified by means of visual attention for the creation of individual photo selections. Users freely viewed photos of a collection without any specific instruction on what to fixate, while their gaze paths were recorded. By comparing gaze-based and baseline photo selections to manually created selections, the worth of eye tracking data in the identification of important photos is shown. In the analysis of the data, the characteristics of gaze data has to be considered, for example, inaccurate and ambiguous data. The aggregation of gaze data, collected from several users, is one suggested approach for dealing with this kind of data.
The results of the performed experiments show the value of gaze data as source of information. It allows to benefit from human abilities where algorithms still have problems to perform satisfyingly.
Through the increasing availability of access to the web, more and more interactions between people take place in online social networks, such as Twitter or Facebook, or sites where opinions can be exchanged. At the same time, knowledge is made openly available for many people, such as by the biggest collaborative encyclopedia Wikipedia and diverse information in Internet forums and on websites. These two kinds of networks - social networks and knowledge networks - are highly dynamic in the sense that the links that contain the important information about the relationships between people or the relations between knowledge items are frequently updated or changed. These changes follow particular structural patterns and characteristics that are far less random than expected.
The goal of this thesis is to predict three characteristic link patterns for the two network types of interest: the addition of new links, the removal of existing links and the presence of latent negative links. First, we show that the prediction of link removal is indeed a new and challenging problem. Even if the sociological literature suggests that reasons for the formation and resolution of ties are often complementary, we show that the two respective prediction problems are not. In particular, we show that the dynamics of new links and unlinks lead to the four link states of growth, decay, stability and instability. For knowledge networks we show that the prediction of link changes greatly benefits from the usage of temporal information; the timestamp of link creation and deletion events improves the prediction of future link changes. For that, we present and evaluate four temporal models that resemble different exploitation strategies. Focusing on directed social networks, we conceptualize and evaluate sociological constructs that explain the formation and dissolution of relationships between users. Measures based on information about past relationships are extremely valuable for predicting the dissolution of social ties. Hence, consistent for knowledge networks and social networks, temporal information in a network greatly improves the prediction quality. Turning again to social networks, we show that negative relationship information such as distrust or enmity can be predicted from positive known relationships in the network. This is particularly interesting in networks where users cannot label their relationships to other users as negative. For this scenario we show how latent negative relationships can be predicted.
This thesis presents novel approaches for integrating context information into probabilistic models. Data from social media is typically associated with metadata, which includes context information such as timestamps, geographical coordinates or links to user profiles. Previous studies showed the benefits of using such context information in probabilistic models, e.g.\ improved predictive performance. In practice, probabilistic models which account for context information still play a minor role in data analysis. There are multiple reasons for this. Existing probabilistic models often are complex, the implementation is difficult, implementations are not publicly available, or the parameter estimation is computationally too expensive for large datasets. Additionally, existing models are typically created for a specific type of content and context and lack the flexibility to be applied to other data.
This thesis addresses these problems by introducing a general approach for modelling multiple, arbitrary context variables in probabilistic models and by providing efficient inference schemes and implementations.
In the first half of this thesis, the importance of context and the potential of context information for probabilistic modelling is shown theoretically and in practical examples. In the second half, the example of topic models is employed for introducing a novel approach to context modelling based on document clusters and adjacency relations in the context space. They can cope with areas of sparse observations and These models allow for the first time the efficient, explicit modelling of arbitrary context variables including cyclic and spherical context (such as temporal cycles or geographical coordinates). Using the novel three-level hierarchical multi-Dirichlet process presented in this thesis, the adjacency of ontext clusters can be exploited and multiple contexts can be modelled and weighted at the same time. Efficient inference schemes are derived which yield interpretable model parameters that allow analyse the relation between observations and context.
The Web contains some extremely valuable information; however, often poor quality, inaccurate, irrelevant or fraudulent information can also be found. With the increasing amount of data available, it is becoming more and more difficult to distinguish truth from speculation on the Web. One of the most, if not the most, important criterion used to evaluate data credibility is the information source, i.e., the data origin. Trust in the information source is a valuable currency users have to evaluate such data. Data popularity, recency (or the time of validity), reliability, or vagueness ascribed to the data may also help users to judge the validity and appropriateness of information sources. We call this knowledge derived from the data the provenance of the data. Provenance is an important aspect of the Web. It is essential in identifying the suitability, veracity, and reliability of information, and in deciding whether information is to be trusted, reused, or even integrated with other information sources. Therefore, models and frameworks for representing, managing, and using provenance in the realm of Semantic Web technologies and applications are critically required. This thesis highlights the benefits of the use of provenance in different Web applications and scenarios. In particular, it presents management frameworks for querying and reasoning in the Semantic Web with provenance, and presents a collection of Semantic Web tools that explore provenance information when ranking and updating caches of Web data. To begin, this thesis discusses a highly exible and generic approach to the treatment of provenance when querying RDF datasets. The approach re-uses existing RDF modeling possibilities in order to represent provenance. It extends SPARQL query processing in such a way that given a SPARQL query for data, one may request provenance without modifying it. The use of provenance within SPARQL queries helps users to understand how RDF facts arederived, i.e., it describes the data and the operations used to produce the derived facts. Turning to more expressive Semantic Web data models, an optimized algorithm for reasoning and debugging OWL ontologies with provenance is presented. Typical reasoning tasks over an expressive Description Logic (e.g., using tableau methods to perform consistency checking, instance checking, satisfiability checking, and so on) are in the worst case doubly exponential, and in practice are often likewise very expensive. With the algorithm described in this thesis, however, one can efficiently reason in OWL ontologies with provenance, i.e., provenance is efficiently combined and propagated within the reasoning process. Users can use the derived provenance information to judge the reliability of inferences and to find errors in the ontology. Next, this thesis tackles the problem of providing to Web users the right content at the right time. The challenge is to efficiently rank a stream of messages based on user preferences. Provenance is used to represent preferences, i.e., the user defines his preferences over the messages' popularity, recency, etc. This information is then aggregated to obtain a joint ranking. The aggregation problem is related to the problem of preference aggregation in Social Choice Theory. The traditional problem formulation of preference aggregation assumes a I fixed set of preference orders and a fixed set of domain elements (e.g. messages). This work, however, investigates how an aggregated preference order has to be updated when the domain is dynamic, i.e., the aggregation approach ranks messages 'on the y' as the message passes through the system. Consequently, this thesis presents computational approaches for online preference aggregation that handle the dynamic setting more efficiently than standard ones. Lastly, this thesis addresses the scenario of caching data from the Linked Open Data (LOD) cloud. Data on the LOD cloud changes frequently and applications relying on that data - by pre-fetching data from the Web and storing local copies of it in a cache - need to continually update their caches. In order to make best use of the resources (e.g., network bandwidth for fetching data, and computation time) available, it is vital to choose a good strategy to know when to fetch data from which data source. A strategy to cope with data changes is to check for provenance. Provenance information delivered by LOD sources can denote when the resource on the Web has been changed last. Linked Data applications can benefit from this piece of information since simply checking on it may help users decide which sources need to be updated. For this purpose, this work describes an investigation of the availability and reliability of provenance information in the Linked Data sources. Another strategy for capturing data changes is to exploit provenance in a time-dependent function. Such a function should measure the frequency of the changes of LOD sources. This work describes, therefore, an approach to the analysis of data dynamics, i.e., the analysis of the change behavior of Linked Data sources over time, followed by the investigation of different scheduling update strategies to keep local LOD caches up-to-date. This thesis aims to prove the importance and benefits of the use of provenance in different Web applications and scenarios. The exibility of the approaches presented, combined with their high scalability, make this thesis a possible building block for the Semantic Web proof layer cake - the layer of provenance knowledge.
Navigation is a natural way to explore and discover content in a digital environment. Hence, providers of online information systems such as Wikipedia---a free online encyclopedia---are interested in providing navigational support to their users. To this end, an essential task approached in this thesis is the analysis and modeling of navigational user behavior in information networks with the goal of paving the way for the improvement and maintenance of web-based systems. Using large-scale log data from Wikipedia, this thesis first studies information access by contrasting search and navigation as the two main information access paradigms on the Web. Second, this thesis validates and builds upon existing navigational hypotheses to introduce an adaptation of the well-known PageRank algorithm. This adaptation is an improvement of the standard PageRank random surfer navigation model that results in a more "reasonable surfer" by accounting for the visual position of links, the information network regions they lead to, and the textual similarity between the link source and target articles. Finally, using agent-based simulations, this thesis compares user models that have a different knowledge of the network topology in order to investigate the amount and type of network topological information needed for efficient navigation. An evaluation of agents' success on four different networks reveals that in order to navigate efficiently, users require only a small amount of high-quality knowledge of the network topology. Aside from the direct benefits to content ranking provided by the "reasonable surfer" version of PageRank, the empirical insights presented in this thesis may also have an impact on system design decisions and Wikipedia editor guidelines, i.e., for link placement and webpage layout.
The distributed setting of RDF stores in the cloud poses many challenges. One such challenge is how the data placement on the compute nodes can be optimized to improve the query performance. To address this challenge, several evaluations in the literature have investigated the effects of existing data placement strategies on the query performance. A common drawback in theses evaluations is that it is unclear whether the observed behaviors were caused by the data placement strategies (if different RDF stores were evaluated as a whole) or reflect the behavior in distributed RDF stores (if cloud processing frameworks like Hadoop MapReduce are used for the evaluation). To overcome these limitations, this thesis develops a novel benchmarking methodology for data placement strategies that uses a data-placement-strategy-independent distributed RDF store to analyze the effect of the data placement strategies on query performance.
With this evaluation methodology the frequently used data placement strategies have been evaluated. This evaluation challenged the commonly held belief that data placement strategies that emphasize local computation, such as minimal edge-cut cover, lead to faster query executions. The results indicate that queries with a high workload may be executed faster on hash-based data placement strategies than on, e.g., minimal edge-cut covers. The analysis of the additional measurements indicates that vertical parallelization (i.e., a well-distributed workload) may be more important than horizontal containment (i.e., minimal data transport) for efficient query processing.
Moreover, to find a data placement strategy with a high vertical parallelization, the thesis tests the hypothesis that collocating small connected triple sets on the same compute node while balancing the amount of triples stored on the different compute nodes leads to a high vertical parallelization. Specifically, the thesis proposes two such data placement strategies. The first strategy called overpartitioned minimal edge-cut cover was found in the literature and the second strategy is the newly developed molecule hash cover. The evaluation revealed a balanced query workload and a high horizontal containment, which lead to a high vertical parallelization. As a result these strategies showed a better query performance than the frequently used data placement strategies.
The Web is an essential component of moving our society to the digital age. We use it for communication, shopping, and doing our work. Most user interaction in the Web happens with Web page interfaces. Thus, the usability and accessibility of Web page interfaces are relevant areas of research to make the Web more useful. Eye tracking is a tool that can be helpful in both areas, performing usability testing and improving accessibility. It can be used to understand users' attention on Web pages and to support usability experts in their decision-making process. Moreover, eye tracking can be used as an input method to control an interface. This is especially useful for people with motor impairment, who cannot use traditional input devices like mouse and keyboard. However, interfaces on Web pages become more and more complex due to dynamics, i.e., changing contents like animated menus and photo carousels. We need general approaches to comprehend dynamics on Web pages, allowing for efficient usability analysis and enjoyable interaction with eye tracking. In the first part of this thesis, we report our work on improving gaze-based analysis of dynamic Web pages. Eye tracking can be used to collect the gaze signals of users, who browse a Web site and its pages. The gaze signals show a usability expert what parts in the Web page interface have been read, glanced at, or skipped. The aggregation of gaze signals allows a usability expert insight into the users' attention on a high-level, before looking into individual behavior. For this, all gaze signals must be aligned to the interface as experienced by the users. However, the user experience is heavily influenced by changing contents, as these may cover a substantial portion of the screen. We delineate unique states in Web page interfaces including changing contents, such that gaze signals from multiple users can be aggregated correctly. In the second part of this thesis, we report our work on improving the gaze-based interaction with dynamic Web pages. Eye tracking can be used to retrieve gaze signals while a user operates a computer. The gaze signals may be interpreted as input controlling an interface. Nowadays, eye tracking as an input method is mostly used to emulate mouse and keyboard functionality, hindering an enjoyable user experience. There exist a few Web browser prototypes that directly interpret gaze signals for control, but they do not work on dynamic Web pages. We have developed a method to extract interaction elements like hyperlinks and text inputs efficiently on Web pages, including changing contents. We adapt the interaction with those elements for eye tracking as the input method, such that a user can conveniently browse the Web hands-free. Both parts of this thesis conclude with user-centered evaluations of our methods, assessing the improvements in the user experience for usability experts and people with motor impairment, respectively.
As a multilingual system,Wikipedia provides many challenges for academics and engineers alike. One such challenge is cultural contextualisation of Wikipedia content, and the lack of approaches to effectively quantify it. Additionally, what seems to lack is the intent of establishing sound computational practices and frameworks for measuring cultural variations in the data. Current approaches seem to mostly be dictated by the data availability, which makes it difficult to apply them in other contexts. Another common drawback is that they rarely scale due to a significant qualitative or translation effort. To address these limitations, this thesis develops and tests two modular quantitative approaches. They are aimed at quantifying culture-related phenomena in systems which rely on multilingual user-generated content. In particular, they allow to: (1) operationalise a custom concept of culture in a system; (2) quantify and compare culture-specific content- or coverage biases in such a system; and (3) map a large scale landscape of shared cultural interests and focal points. Empirical validation of these approaches is split into two parts. First, an approach to mapping Wikipedia communities of shared co-editing interests is validated on two large Wikipedia datasets comprising multilateral geopolitical and linguistic editor communities. Both datasets reveal measurable clusters of consistent co-editing interest, and computationally confirm that these clusters correspond to existing colonial, religious, socio economic, and geographical ties. Second, an approach to quantifying content differences is validated on a multilingual Wikipedia dataset, and a multi-platform (Wikipedia and Encyclopedia Britannica) dataset. Both are limited to a selected knowledge domain of national history. This analysis allows, for the first time on the large scale, to quantify and visualise the distribution of historical focal points in the articles on national histories. All results are cross-validated either by domain experts, or external datasets.
Main thesis contributions. This thesis: (1) presents an effort to formalise the process of measuring cultural variations in user-generated data; (2) introduces and tests two novel approaches to quantifying cultural contextualisation in multilingual data; (3) synthesises a valuable overview of literature on defining and quantifying culture; (4) provides important empirical insights on the effect of culture on Wikipedia content and coverage; demonstrates that Wikipedia is not contextfree, and these differences should not be treated as noise, but rather, as an important feature of the data. (5) makes practical service contributions through sharing data and visualisations.
Graph-based data formats are flexible in representing data. In particular semantic data models, where the schema is part of the data, gained traction and commercial success in recent years. Semantic data models are also the basis for the Semantic Web - a Web of data governed by open standards in which computer programs can freely access the provided data. This thesis is concerned with the correctness of programs that access semantic data. While the flexibility of semantic data models is one of their biggest strengths, it can easily lead to programmers accidentally not accounting for unintuitive edge cases. Often, such exceptions surface during program execution as run-time errors or unintended side-effects. Depending on the exact condition, a program may run for a long time before the error occurs and the program crashes.
This thesis defines type systems that can detect and avoid such run-time errors based on schema languages available for the Semantic Web. In particular, this thesis uses the Web Ontology Language (OWL) and its theoretic underpinnings, i.e., description logics, as well as the Shapes Constraint Language (SHACL) to define type systems that provide type-safe data access to semantic data graphs. Providing a safe type system is an established methodology for proving the absence of run-time errors in programs without requiring execution. Both schema languages are based on possible world semantics but differ in the treatment of incomplete knowledge. While OWL allows for modelling incomplete knowledge through an open-world semantics, SHACL relies on a fixed domain and closed-world semantics. We provide the formal underpinnings for type systems based on each of the two schema languages. In particular, we base our notion of types on sets of values which allows us to specify a subtype relation based on subset semantics. In case of description logics, subsumption is a routine problem. For
the type system based on SHACL, we are able to translate it into a description
logic subsumption problem.