Refine
Document Type
- Doctoral thesis (5)
Has Fulltext
- yes (5)
Is part of the Bibliography
- no (5)
Language
- English (5)
Keywords
- Abstammungsformel (1)
- Anfrageauswertung (1)
- Anfrageoptimierung (1)
- Anomalieerkennung (1)
- Anomaly Detection (1)
- CQQL (1)
- Cluster-Analyse (1)
- Clustering distance (1)
- Clustering sozialer Netzwerke (1)
- Clustering-Abstand (1)
Institute
Mobile cyber-physical systems (MCPSs) such as motor vehicles, railed vehicles, aircraft, or spacecraft are commonly used in our life today. These systems are location-independent and embedded in a physical environment which is usually harsh and uncertain. MCPSs are equipped with a wide range of sensors that continuously produce sensor data streams. It is mandatory to process these data streams in an appropriate manner in order to satisfy different monitoring objectives, and it is anticipated that the complexity of MCPSs will continue to increase in the future. For instance, this includes the system description and the amount of data that must be processed. Accordingly, it is necessary to monitor these systems in order to provide reliability and to avoid critical damage. Monitoring is usually a semi-automatic process while human experts are responsible for consequent decisions. Thus, appropriate monitoring approaches are required to both provide a reasonably precise monitoring process and to reduce the complexity of the monitoring process itself.
The contribution of the present thesis is threefold. First, a knowledge discovery cycle (KDC) has been developed, which aims to combine the research areas of knowledge discovery in databases and knowledge discovery from data streams to monitor MCPSs. The KDC is a cyclic process chain comprising an online subcycle and an offline subcycle. Second, a new data stream anomaly detection algorithm has been developed. Since it is necessary to identify a large number of system states automatically during operation, data stream anomaly detection becomes a key task for monitoring MCPSs. Third, the KDC and the anomaly detection algorithm have been prototypically implemented and a case study has been performed in a real world scenario relating to the ISS Colombus module.
People in a social network are connected, and their homogeneity is reflected by the similarity of their attributes. For effective clustering, the similarities among people within a cluster must be much higher than the similarities between different clusters. Traditional clustering algorithms like hierarchical (agglomerative) or k-medoids take distances between objects as input and find clusters of objects. The distance functions used should comply with the triangle inequality (TI) property, but sometimes this property may be violated, thus negatively impacting the quality of the generated clusters.
However, in social networks, meaningful clusters can be found even though TI violates. One possibility is a quantum-logic-based clique-guided non-TI clustering approach. The commuting quantum query language (CQQL) is the base for this approach. The CQQL allows the formulation of queries that incorporate both Boolean and similarity conditions. It calculates the similarity value between two objects.
Furthermore, attributes may not have equal impact on similarity and affect the resulting clusters. CQQL incorporates weights to express the varying importance of sub-conditions in a query while preserving consistency with Boolean algebra. This enables personalization of results through relevance feedback (RF).
The main challenge of comparing clusterings is that there is no ground truth data. In such situations, human-generated gold standard clustering can be used. The question is how to compare the performance of clusterings. A noteworthy technique involves counting the pairs of objects that are grouped identically in both clusterings. By doing so, a clustering distance is calculated that measures the dissimilarity.
To validate the non-TI clustering approach, experiments are conducted on social networks of different sizes. Three central questions are addressed by the experiments mentioned: first, is it possible to find meaningful clusters even though TI violates; second, how does a user interact with the system to provide feedback based on their needs; and third, how fast do the detected clusters based on the proposed approach converge to the ideal solution?
To sum up, the experiments’ objective is to demonstrate the validity of a theoretical approach. The research findings presented here provide sufficient evidence for detecting meaningful clusters based on user interaction. Furthermore, the experiments clearly demonstrate that the non-TI clustering approach can be used as an RF technique in clustering.
The search for textual information, e.g., in the form of webpages, is a typical task in modern business and private life. From a user's point of view, the commonly used systems have matured and established common interaction design patterns such as the textual input box that starts virtually every directed search process.
In comparison, the search for multimedia documents (e.g., images or videos) is still in its early years. In other words, a pre-dominant search strategy has not yet evolved. That is, directed and exploratory search approaches fight for user acceptance.
One further discriminative factor of multimedia information retrieval (MMIR) from traditional text-based information retrieval (IR) is that multimedia documents are not necessarily stored with the help of the same data access paradigm.
From a technical point of view, the use of different data access paradigms complicates the retrieval from such collections because the utilized retrieval model has to support these paradigms.
As a consequence, the main challenges in MMIR - the retrieval engine and the user interaction -- have to be addressed in a holistic way. A holistic theoretic perspective on MMIR/IR research is taken by principle of polyrepresentation (PoP), which forms one half of the theoretic background of this dissertation aiming at the development of a preference-based approach to interactive MMIR. Roughly speaking, the PoP theorizes that representations describing a document are based on various cognitive processes dealing with it, e.g., a title, its color or shape features, its creator, or its date of creation. This multitude of representations can be fused to form a conjunctive cognitive overlap (CO) in which highly relevant documents are likely to be contained. This explicit recommendation discriminates the PoP from typical feature fusion approaches often used in MMIR.
However, the PoP does not answer how a retrieval model has to be implemented in a technical sense which is of interest in the field of computer science. One possibility to implement the PoP are quantum mechanics-inspired IR models such as the commuting quantum query language (CQQL) which is used in this thesis.
CQQL is particularly interesting because it integrates data access paradigms used in the fields of DB and IR. In order to respect the dynamic nature of the search process and information need (IN), CQQL allows the personalization of retrieval results using a preference-based relevance feedback (RF) approach called PrefCQQL, which relies on machine-based learning.
Unique features of the PrefCQQL approach range from the support of negative query-by-example (QBE) documents at query formulation time as well as during the interactive retrieval process to the formulation of weak preferences between result documents to express gradual levels of relevance. In addition, inductive preferences can be used from query formulation time onward to learn new CQQL queries.
In order to evaluate the presented polyrepresentative PrefCQQL approach, two kinds of experiments are conducted: a Cranfield-inspired evaluation of CQQL/PrefCQQL's retrieval effectiveness, which is extended by the utilization of user simulations to better fit the requirements of the evaluation of an adaptive IR system, and a usability study that examines three alternative MMIR system UI prototypes. In order to increase the reproducibility and confirmability of the experiments, the source code to all used programs is made available as a supplement to this dissertation.
The mentioned experiments aim at answering two central questions: first, whether the hypotheses of the PoP can be verified in MMIR, and second, whether a usable interactive MMIR system can be built on the basis of the PoP and PrefCQQL?
To answer the first question, different matching functions that partly follow the recommendations of the PoP are evaluated with six different test collections in both an non-interactive and interactive QBE scenario. The results of this experiment are ambivalent.
In non-interactive MMIR, the experimental data does not provide sufficient justification for the statement that PoP-based matching functions will always surpass single features or other matching functions. For instance, the arithmetic mean, which calculates the average similarity between a query's representations and the documents' representations in the collection, surpasses the conjunction and hence the CO of multiple representations in terms of retrieval effectiveness. Nevertheless, the matching function following the PoP is effectiveness stabler than the best performing single representations per collection. Hence, the CO's retrieval performance is more reliable than the usage of single representations.
In contrast, the predictions of the PoP can be verified in the investigated PrefCQQL-based interactive MMIR scenario. However, it is important to note that also the number of available representations has an impact on the retrieval outcome. That is, if too few representations are present in a matching function, the corresponding IN model in PrefCQQL obviously becomes subject to underfitting eventually lowering its retrieval effectiveness. Unfortunately, when the point of sufficient representations to support PrefCQQL is reached could not be revealed in this dissertation.
The second question is answered with the help of a prototypical MMIR system: the Pythia system, which serves both as proof of concept of the CQQL and the PrefCQQL approach. Furthermore, the system supports different information seeking strategies and a seamless transition between them in order to support users with different kinds of IN.
In contrast to traditional data applications, many real-world scenarios nowadays depend on managing and querying huge volumes of uncertain and incomplete data. This new type of applications emerge, for example, when we integrate data from various sources, analyse social/biological/chemical networks or conduct privacy-preserving data mining.
A very promising concept addressing this new kind of probabilistic data applications has been proposed in the form of probabilistic databases. Here, a tuple only belongs to its table or query answer with a specific likelihood. That probability expresses the uncertainty about the given data or the confidence in the answer. The most challenging task for probabilistic databases is query evaluation. In fact, there are even simple relational queries for which determining the occurrence probability of a single answer tuple is hard for #P.
Lineage formulas constitute the central concept under investigation in this work. In short, the mechanism behind lineage formulas facilitates the representation and evaluation of events of the probability space, which is defined by a probabilistic database. On the basis of lineage formulas, we devise a framework that is designed as a combination of a relational database layer and an additional probabilistic query engine.
In particular, the following three aspects are studied:
(i) an efficient construction of lineage formulas,
(ii) an orthogonal combination of lineage optimization techniques, which are performed within the relational database layer and the probabilistic query engine, and
(iii) effective and compact data structures to represent lineage formulas within a probabilistic query engine.
The developed framework provides a novel lineage construction method that is able to construct nested lineage formulas, to avoid large tuple sets within the relational database layer tuples, and to provide full relational algebra support. In addition, the proposed system completely resolves the conflict between the contradicting query plans optimized for the relational database layer and the probabilistic query engine.
The main part of the research outlined in this thesis is to develop Deep Learning models for the linguistic interpretation of the visual contents. This part is split into two research problems: interactive region segmentation and captioning, and selective texture labeling. In the first attempt, we proposed a novel hybrid Deep Learning architecture whereby the user is able to specify an arbitrary region of the image that should be highlighted and described. The proposed model alternates the bounding box indications of the standard object localization process with the output of a deep interactive segmentation module to achieve a better understanding of the dense image captioning and improve the object localization accuracy. The idea of the next part is to establish a bidirectional correlation between deep texture representation and its linguistic description via a hybrid CNN-RNN model that enables end-to-end learning of the selective texture labeling. This novel architecture provides new opportunities to describe, search, and also retrieve texture images from their linguistic descriptions. To be able to train such a model, we generated a multi-label texture dataset that covers color, material, and pattern labeling simultaneously. Our contribution to the automatic generation of texture descriptions provides an excellent opportunity to enrich the existing vocabulary of the image captioning. Such a conceptual extension can be used for fine-grained captioning applicable in geology, meteorology and other natural sciences where fine-grained image structures are of importance to understand complicated patterns. Apart from Deep Learning technologies, in the final section of the thesis, we proposed a novel approach to define mathematical morphology on color images. To this end, we converted common RGB-values of the color images into a new biconal color space and then combined two approaches of mathematical morphology to give meaning to the maximum and the minimum of the matrix field data and formulate our novel strategy.