xxx

Research on recommender systems has gained a tremendous popularity in recent years. Although various recommender approaches are available nowadays, there is still a lack of work that tackles real-time recommendation on large and sparse data. To tackle the data sparsity problem, this thesis analyzes different trust-based approaches which improve the accuracy of the usually used Collaborative Filtering recommendation approaches. To show how the trust-based approaches can also be applied to generate real-time recommendations, this thesis extended ScaR, a scalable recommendation framework, with recommendation approaches which calculate the trust values between users using the Apache Solr search engine. Experimental results showed that using trust-based approaches, high quality recommendations can be served in realtime.

Everyone knows the annoying situation when personal items of appreciated value are disappeared and several precious minutes, or even hours, are was- ted for frantic searching. Modern technologies are useful for assisting in such moments. For example electronic key finders triggered by whistling ore remo- te controls are available over years, but the acceptance for these gadgets are rather low. An important field of research on that topic is on using modern smartphones for locating your everyday objects. Today’s smartphones are equipped with various hardware that can be used for retrieving locations. The primary used technology for this purpose is the Global Positioning System (GPS). For example Apple is successfully offering the service Find my iPhone, which can locate a misplaced iPhone with the usage of GPS. The limitation of GPS is the lack of accuracy in urban areas and especially inside of buildings. To counteract this limitation, GPS is often used together with WiFi triangulation, which needs a well developed WiFi infrastructure for proper operation, which is difficult to achieve in private households. The goal of this thesis is to develop an easy to use application for indivi- duals for retrieving their lost items in- and outdoors only with technologies present in their smartphones. A hybrid solution of localization and motion sensing will be used for tracing the user’s location. The focus will be on indoor tracing using accelerometer, gyroscope and compass data. The proto- type is implemented as an iPhone application to record motion and location data and a web application to calculate and visualize the user’s trace. The web application will also provide a user interface for backtracking the user’s trace to a lost item by time filtering or by tagging items. 

During a typical day, we have several social interactions with different people belonging to different semantic groups (e.g. Friends, family, co-workers). In this paper we try to find promising hypothesis to link data collected from a mobile sensing application running on the users smartphone to the social interactions he has during a typical day. We will search for possibilities to reliably determine (1) the number of interactions he has during the day, (2) the length of social interactions, (3) the number of participants, (4) who the participants were and (5) the semantic context of the interaction using data collected by a pilot study, where, additionally to the date collected by the framework, users label their interactions during the day. 

With this thesis we try to determine the feasibility of detecting face-to-face social interactions based on standard smartphone sensors like Bluetooth, Global Positioning System (GPS) data, microphone or magnetic field sen- sor. We try to detect the number of social interactions by leveraging Mobile Sens- ing on modern smartphones. Mobile Sensing is the use of smartphones as ubiquitous sensing devices to collect data. Our focus lies on the standard smartphone sensors provided by the Android Software Development Kit (SDK) as opposed to previous work which mostly leverages only audio sig- nal processing or Bluetooth data. To mine data and collect ground truth data, we write an Android 2 app that collects sensor data using the Funf Open Sensing Framework[1] and addi- tionally allows the user to label their social interaction as they take place. With the app we perform two user studies over the course of three days with three participants each. We collect the data and add additional meta-data for every user during an interview. This meta-data consists of semantic labels for location data and the distinction of social interactions into private and business social interactions. We collected a total of 16M data points for the first group and 35M data points for the second group. Using the collected data and the ground truth labels collected by our partici- pants, we then explore how time of day, audio data, calendar appointments, magnetic field values, Bluetooth data and location data interacts with the number of social interactions of a person. We perform this exploration by creating various visualization for the data points and use time correlation to determine if they influence the social interaction behavior. We find that only calendar appointments provide some correlation with the social interactions and could be used in a detection algorithm to boost the accuracy of the result. The other data points show no correlation during our exploratory evaluation of the collected data. We also find that visualizing the interactions in the form of a heatmap on a map is a visualization that most participants find very interesting. Our participants also made clear that la- beling all social interactions over the course of a day is a very tedious task. We recommend that further research has to include audio signal process- ing and a carefully designed study setup. This design has to include what data needs to be sampled at what frequency and accuracy and must provide further assistance to the user for labeling the data. We release the data mining app and the code used to analyze the data as open source under the MIT License.  

Many people face the problem of misplaced personal items in their daily routine, especially when they are in a hurry, and often waste a lot of time searching these items. There are different gadgets and applications available on the market, which are trying to help people find lost items. Most often, help is given by creating an infrastructure that can locate lost items. This thesis presents a novel approach for finding lost items, namely by helping people re-trace their movements throughout the day. Movements are logged by indoor localization based on mobile phone sensing. An external infrastructure is not needed. The application is based on a step based pedestrian dead reckoning system, which is developed to collect real-time localization data. This data is used to draw a live visualization of the whole trace the user has covered, from where the user can retrieve the position of the lost personal items, after they were tagged using simple speech commands. The results from the field experiment, that was performed with twelve participants of different age and gender, showed that the application could successfully visualize the covered route of the pedestrians and reveal the position of the placed items.  

The amount of multimedia content being created is growing tremendously. In addition, the number of applications for processing, consuming, and sharing multimedia content is growing. Being able to create and process metadata describing this content is an important prerequisite to ensure a correct workflow of applications. The MPEG-7 standard enables the description of different types of multimedia content by creating standardized metadata descriptions. When using MPEG-7 practically, two major drawbacks are identified, namely complexity and fuzziness. Complexity is mainly based on the comprehensiveness of MPEG-7, while fuzziness is a result of the syntax variability. The notion of MPEG-7 profiles were introduced in order to address and possibly solve these issues. A profile defines the usage and semantics of MPEG-7 tailored to a particular application domain. Thus usage instructions and explanations, denoted as semantic constraints, can be expressed as English prose. However, this textual explanations leave space for potential misinterpretations since they have no formal grounding. While checking the conformance of an MPEG-7 profile description is possible on a syntactical level, the semantic constraints currently cannot be checked in an automated way. Being unable to handle the semantic constraints, inconsistent MPEG-7 profile descriptions can be created or processed leading to potential interoperability issues. Thus an approach for formalizing the semantic constraints of MPEG-7 profiles using ontologies and logical rules is presented in this thesis. Ontologies are used to model the characteristics of the different profiles with respect to the semantic constraints, while validation rules detect and flag violations of these constraints. In similar manner, profile-independent temporal semantic constraints are also formalized. The presented approach is the basis for a semantic validation service for MPEG-7 profile descriptions, called VAMP. VAMP verifies the conformance of a given MPEG-7 profile description with a selected MPEG-7 profile specification in terms of syntax and semantics. Three different profiles are integrated in VAMP. The temporal semantic constraints are also considered. As a proof of concept, VAMP is implemented as a web application for human users and as a RESTful web service for software agents.  

The goal of this thesis is to improve query suggestions for rare queries on faceted documents. While there has been extensive work on query suggestions for single facet documents there is only little known about how to provide query suggestions in the context of faceted documents. The constraint to provide suggestions also for uncommon or even previously unseen queries (so-called rare queries) increases the difficulty of the problem as the commonly used technique of mining query logs can not be easily applied.

In this thesis it was further assumed that the user of the information retrieval system always searches for one specific document - leading to uniformly distributed queries. Under these constraints it was tried to exploit the structure of the faceted documents to provide helpful query suggestions. In addition to theoretical exploration of such improvements a custom datastructure was developed to efficiently provide interactive query suggestions. Evaluation of the developed query suggestion algorithms was done on multiple document collections by comparing them to a baseline algorithm that reduces faceted documents to single facet documents. Results are promising as the final version of the new query suggestion algorithm consistently outperformed the baseline.

Motivation for and potential application of this work can be found in call centers for customer support. For call center employees it is crucial to quickly locate relevant customer information - information that is available in structured form (and can thus easily be transformed into faceted documents).

“Wiktionary”, is a free dictionary which is part of Wikmedia Foundation. This webpage contains translations, etymologies, synonyms and pronunciations of words in multiple languages in that case we just focus on English.

A syntactic analyser (parser) turns the entry text in other structures, which will make easier the analysis and capture of nest entrance.

Unter Wissenschaftlern ist Twitter ein sehr beliebtes soziales Netzwerk. Dort diskutieren sie verschiedenste Themen und werben für neue Ideen oder präsentieren Ergebnisse ihrer aktuellen Forschungsarbeit. Die in dieser Arbeit durchgeführten Experimente beruhen auf einem Twitter-Datensatz welcher aus den Tweets von Informatikern, deren Forschungsbereiche bekannt sind, besteht. Die vorliegende Diplomarbeit kann grob in vier Teile unterteilt werden: Zunächst wird beschrieben, wie der Twitter-Datensatz erstellt wurde. Danach werden diverse Statistiken zu diesem Datensatz präsentiert. Beispielsweise wurden die meisten Tweets während der Arbeitszeit erstellt und die Nutzer sind unterschiedlich stark aktiv. Aus den Follower-Beziehungen der Nutzer wurde ein Netzwerk erstellt, welches nachweislich small world Eigenschaften hat. Darüber hinaus sind in diesem Netzwerk auch die verschiedenen Forschungsbereiche sichtbar. Der dritte Teil dieser Arbeit ist der Untersuchung der Hashtagbenutzung gewidmet. Dabei zeigte sich, dass die meisten Hashtags nur selten benutzt werden. Über den gesamten Beobachtungszeitraum betrachtet ändert sich die Verwendung von Hashtags kaum, jedoch gibt es viele kurzfristige Schwankungen. Da die Forschungsbereiche der Nutzer bekannt sind, können auch die Bereiche der Hashtags bestimmt werden. Dadurch können die Hashtags dann in fachspezifische und generelle Hashtags unterteilt werden. Die Analyse der Weitergabe von Hashtags über das Twitter-Netzwerk wird im vierten Teil mittels sogenannter Informationsflussbäume betrachtet. Aufgrund dieser Informationsflussbäume kann gemessen werden wie gut ein Nutzer Informationen verbreitet und erzeugt. Dabei wurde auch die Hypothese bestätigt, dass diese Eigenschaften von der Anzahl der Tweets und Retweets und der Stellung im sozialen Netzwerk abhängen. Jedoch ist dieser Zusammenhang nur in Einzelfällen stark ausgeprägt.