Adaptive Similarity Search

An important aspect when looking for similar data objects is the underlying concept of similarity. Since this concept might depend on the current user or even the current situation it is important that a similarity search system is capable to adapt to multiple similarity functions.

Objectives

The goal of this project is the development of efficient similarity search systems for various types of multimedia objects. The systems will employ multiple similarity measures retrieving different objects depending in the current user profile or workflow step.

Tasks

Capturing User Intention

To train a function using data mining and machine learning techniques it is necessary to have examples for the given user preferences. However, manually labelling a large set of object comparisons is an annoying and timing consumuing task. Thus, in order to capture the user preferences, we develop methods that monitore user behaviour and thus, automatically derive training data.

Learning Similarity Functions

Learning a similariy function is the core of an adaptiv search system. The result if a function that compares objects under the aspects considered to be most important to the given user. In this task, we train multi-represented simialrity measures depending on serveral feature descriptions and general similarity mappings going beyond the framework of the well-known Mahalanobis distance. A further aspect is the problem of transferring a function being learned on the preferences of user to a new unknown user.

Indexing for Multiple Similarity Functions

A final task of this project is the development of indexing techniques for the similarity funcitons learned in the previous task. A key aspect of this problem is the trade-off between building up an index for each similarity function or building up a more general data structure supporting multiple functions.

Publications

• Kriegel H.-P., Kunath P., Pryakhin A., Schubert M.: MUSE: Multi-Represented Similarity Estimation in proc. 24th International Conference on Data Engineering (ICDE 2008), Cancún, México, 2008
• Kriegel H.-P., Kunath P., Pryakhin A., Schubert M.: Distribution-Based Similarity for Multi-Represented Multimedia Objects in proc. 14th international Multimedia Modeling Conference (MMM 2008), Kyoto, Japan, 2008

Team