Formerprojects
From Matthias_Schubert
Former projects
Shared-E-Fleet
In this project, we develop routing and spatio-temporal prediction solutions for shared car-pools of electronic cars. Project Website: Shared-E-Fleet. (funded by the German Ministry of Economics and Technology (BMWi))
Data Mining in Multiplayer Online Games
Multiplayer online games develop more and more into major area in the entertainment industrie. In this project, we examine the use of data mining techniques for better dynamic resource allocation. Another major area of data mining in online game is dectecting player which are not behaving due to rules of the game. The Loot Box
Adaptive Image Retrieval
Similarity search in image datbases is a well known problem which has been studied in the areas of imaging, pattern recognition and databases. An important aspect of this task is to capture what types of images are really considered to be similar by a human user. Thus, as part of this project I am interested in learning similarity functions adapting to the user idea of similarity. Another important part of this project will be to find new methods to capture what kind of objects a certain user considers to be similar.
Indexing Compound Object Descriptions
In many application, there are heterogenous data describing an object. For example, images can be represented by sets of local descriptors, global descriptors, tags or surrounding text. For efficient similarity search, it is now important to develope indexing methods that can employ similarity measures combining a variety of the available aspects.
Multi-Preference Path Planing
Navigation systems and other spatial applications rely on complex maps that store a variety of informaitons about streets and crossings. This project is dedicated to find proper routes through a given spatial network under consideration of multiple criteria.
Traffic Mining
The availability of dynamic information about the traffic in a given road network is constantly growing. In this project, we develop data mining methods that try to predict traffic incidents and approximate traffic density based on the available information about recent traffic.
Website Mining
In this project, we examined problems that are concerned with complete websites instead of single HTML documents. A web site consists of the subgraph of the WWW which is dedicated to represent the same organization or has a common purpose. To assign a website to a certain class like online shop, we developed several classification methods and examined the use of the linkage. Furthmore, we developed a focused crawler that aims at efficiently retrieving large numbers of websites having a certain purpose. For example, finding community websites in order to build up a database of upcoming construction projects.
Indexing Uncertain Objects with the Gauss-Tree
The Gaussian or normal distribution is one of the most common probability distributions for multiple purposes. Thus, we developed an index structure called the Gauss-Tree which is capable to efficiently manage large numbers of Gaussian. The Gauss-Tree was shown to support query with different probability models and can be extend to manage arbitrary distribution functions by using an approximation with Gaussian mixture models.
Multirepresented Data Mining
It is quite common that there exist multiple ways of representing the same real-world object in a feature vector. A combination of these feature representation often yields a significant quality increase of the results of clustering and classificaiton algorithms. In this project, we developed methods for multi-represented density-based clustering. Furthermore, we proposed several classification methods that dynamicalley combine the classification results in each representation depending on the currently classified object.
Multi-Instance Data Mining
It often occures that objects are described as sets of instances instead of a sinlge feature vector. Most of the standard data mining methods are extendable to this kind of object representation in an straight-forward way. In this projecte, we proposd two concept-based clustering mehtods for multi-instance objecs which implicitely assume that each instance belongs to a latend concept in the underlying feature space.
