@phdthesis{Gruetze2018,
  author    = {Gr{\"u}tze, Toni},
  title     = {Adding value to text with user-generated content},
  school      = {Universit{\"a}t Potsdam},
  pages     = {ii, 114},
  year      = {2018},
  abstract  = {In recent years, the ever-growing amount of documents on the Web as well as in closed systems for private or business contexts led to a considerable increase of valuable textual information about topics, events, and entities. It is a truism that the majority of information (i.e., business-relevant data) is only available in unstructured textual form. The text mining research field comprises various practice areas that have the common goal of harvesting high-quality information from textual data. These information help addressing users' information needs. In this thesis, we utilize the knowledge represented in user-generated content (UGC) originating from various social media services to improve text mining results. These social media platforms provide a plethora of information with varying focuses. In many cases, an essential feature of such platforms is to share relevant content with a peer group. Thus, the data exchanged in these communities tend to be focused on the interests of the user base. The popularity of social media services is growing continuously and the inherent knowledge is available to be utilized. We show that this knowledge can be used for three different tasks. Initially, we demonstrate that when searching persons with ambiguous names, the information from Wikipedia can be bootstrapped to group web search results according to the individuals occurring in the documents. We introduce two models and different means to handle persons missing in the UGC source. We show that the proposed approaches outperform traditional algorithms for search result clustering. Secondly, we discuss how the categorization of texts according to continuously changing community-generated folksonomies helps users to identify new information related to their interests. We specifically target temporal changes in the UGC and show how they influence the quality of different tag recommendation approaches. Finally, we introduce an algorithm to attempt the entity linking problem, a necessity for harvesting entity knowledge from large text collections. The goal is the linkage of mentions within the documents with their real-world entities. A major focus lies on the efficient derivation of coherent links. For each of the contributions, we provide a wide range of experiments on various text corpora as well as different sources of UGC. The evaluation shows the added value that the usage of these sources provides and confirms the appropriateness of leveraging user-generated content to serve different information needs.},
  language  = {en}
}
@phdthesis{Amirkhanyan2019,
  author    = {Amirkhanyan, Aragats},
  title     = {Methods and frameworks for GeoSpatioTemporal data analytics},
  doi       = {10.25932/publishup-44168},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-441685},
  school      = {Universit{\"a}t Potsdam},
  pages     = {xxiv, 133},
  year      = {2019},
  abstract  = {In the era of social networks, internet of things and location-based services, many online services produce a huge amount of data that have valuable objective information, such as geographic coordinates and date time. These characteristics (parameters) in the combination with a textual parameter bring the challenge for the discovery of geospatiotemporal knowledge. This challenge requires efficient methods for clustering and pattern mining in spatial, temporal and textual spaces. In this thesis, we address the challenge of providing methods and frameworks for geospatiotemporal data analytics. As an initial step, we address the challenges of geospatial data processing: data gathering, normalization, geolocation, and storage. That initial step is the basement to tackle the next challenge -- geospatial clustering challenge. The first step of this challenge is to design the method for online clustering of georeferenced data. This algorithm can be used as a server-side clustering algorithm for online maps that visualize massive georeferenced data. As the second step, we develop the extension of this method that considers, additionally, the temporal aspect of data. For that, we propose the density and intensity-based geospatiotemporal clustering algorithm with fixed distance and time radius. Each version of the clustering algorithm has its own use case that we show in the thesis. In the next chapter of the thesis, we look at the spatiotemporal analytics from the perspective of the sequential rule mining challenge. We design and implement the framework that transfers data into textual geospatiotemporal data - data that contain geographic coordinates, time and textual parameters. By this way, we address the challenge of applying pattern/rule mining algorithms in geospatiotemporal space. As the applicable use case study, we propose spatiotemporal crime analytics -- discovery spatiotemporal patterns of crimes in publicly available crime data. The second part of the thesis, we dedicate to the application part and use case studies. We design and implement the application that uses the proposed clustering algorithms to discover knowledge in data. Jointly with the application, we propose the use case studies for analysis of georeferenced data in terms of situational and public safety awareness.},
  language  = {en}
}
@phdthesis{Haider2013,
  author    = {Haider, Peter},
  title     = {Prediction with Mixture Models},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus-69617},
  school      = {Universit{\"a}t Potsdam},
  year      = {2013},
  abstract  = {Learning a model for the relationship between the attributes and the annotated labels of data examples serves two purposes. Firstly, it enables the prediction of the label for examples without annotation. Secondly, the parameters of the model can provide useful insights into the structure of the data. If the data has an inherent partitioned structure, it is natural to mirror this structure in the model. Such mixture models predict by combining the individual predictions generated by the mixture components which correspond to the partitions in the data. Often the partitioned structure is latent, and has to be inferred when learning the mixture model. Directly evaluating the accuracy of the inferred partition structure is, in many cases, impossible because the ground truth cannot be obtained for comparison. However it can be assessed indirectly by measuring the prediction accuracy of the mixture model that arises from it. This thesis addresses the interplay between the improvement of predictive accuracy by uncovering latent cluster structure in data, and further addresses the validation of the estimated structure by measuring the accuracy of the resulting predictive model. In the application of filtering unsolicited emails, the emails in the training set are latently clustered into advertisement campaigns. Uncovering this latent structure allows filtering of future emails with very low false positive rates. In order to model the cluster structure, a Bayesian clustering model for dependent binary features is developed in this thesis. Knowing the clustering of emails into campaigns can also aid in uncovering which emails have been sent on behalf of the same network of captured hosts, so-called botnets. This association of emails to networks is another layer of latent clustering. Uncovering this latent structure allows service providers to further increase the accuracy of email filtering and to effectively defend against distributed denial-of-service attacks. To this end, a discriminative clustering model is derived in this thesis that is based on the graph of observed emails. The partitionings inferred using this model are evaluated through their capacity to predict the campaigns of new emails. Furthermore, when classifying the content of emails, statistical information about the sending server can be valuable. Learning a model that is able to make use of it requires training data that includes server statistics. In order to also use training data where the server statistics are missing, a model that is a mixture over potentially all substitutions thereof is developed. Another application is to predict the navigation behavior of the users of a website. Here, there is no a priori partitioning of the users into clusters, but to understand different usage scenarios and design different layouts for them, imposing a partitioning is necessary. The presented approach simultaneously optimizes the discriminative as well as the predictive power of the clusters. Each model is evaluated on real-world data and compared to baseline methods. The results show that explicitly modeling the assumptions about the latent cluster structure leads to improved predictions compared to the baselines. It is beneficial to incorporate a small number of hyperparameters that can be tuned to yield the best predictions in cases where the prediction accuracy can not be optimized directly.},
  language  = {en}
}