Marc Wichterich

Efficient EMD-based similarity search in multimedia databases via flexible dimensionality reduction

The Earth Movers Distance (EMD) was developed in computer vision as a flexible similarity model that utilizes similarities in feature space to define a high quality similarity measure in feature representation space. It has been successfully adopted in a multitude of applications with low to medium dimensionality. However, multimedia applications commonly exhibit high-dimensional feature representations for which the computational complexity of the EMD hinders its adoption. An efficient query processing approach that mitigates and overcomes this effect is crucial. We propose novel dimensionality reduction techniques for the EMD in a filter-and-refine architecture for efficient lossless retrieval. Thorough experimental evaluation on real world data sets demonstrates a substantial reduction of the number of expensive high-dimensional EMD computations and thus remarkably faster response times. Our techniques are fully flexible in the number of reduced dimensions, which is a novel feature in approximation techniques for the EMD.

Anticipatory DTW for efficient similarity search in time series databases

Time series arise in many different applications in the form of sensor data, stocks data, videos, and other time-related information. Analysis of this data typically requires searching for similar time series in a database. Dynamic Time Warping (DTW) is a widely used high-quality distance measure for time series. As DTW is computationally expensive, efficient algorithms for fast computation are crucial. In this paper, we propose a novel filter-and-refine DTW algorithm called Anticipatory DTW. Existing algorithms aim at efficiently finding similar time series by filtering the database and computing the DTW in the refinement step. Unlike these algorithms, our approach exploits previously unused information from the filter step during the refinement, allowing for faster rejection of false candidates. We characterize a class of applicable filters for our approach, which comprises state-of-the-art lower bounds of the DTW. Our novel anticipatory pruning incurs hardly any over-head and no false dismissals. We demonstrate substantial efficiency improvements in thorough experiments on synthetic and real world time series databases and show that our technique is highly scalable to multivariate, long time series and wide DTW bands.

ClasSi: measuring ranking quality in the presence of object classes with similarity information

The quality of rankings can be evaluated by computing their correlation to an optimal ranking. State of the art ranking correlation coefficients like Kendalls τ and Spearmans ρ do not allow for the user to specify similarities between differing object classes and thus treat the transposition of objects from similar classes the same way as that of objects from dissimilar classes. We propose ClasSi, a new ranking correlation coefficient which deals with class label rankings and employs a class distance function to model the similarities between the classes. We also introduce a graphical representation of ClasSi akin to the ROCcurve which describes how the correlation evolves throughout the ranking.

The ClasSi coefficient for the evaluation of ranking quality in the presence of class similarities

Evaluation measures play an important role in the design of new approaches, and often quality is measured by assessing the relevance of the obtained result set. While many evaluation measures based on precision/recall are based on a binary relevance model, ranking correlation coefficients are better suited for multi-class problems. State-of-the-art ranking correlation coefficients like Kendall’s τ and Spearman’s ρ do not allow the user to specify similarities between differing object classes and thus treat the transposition of objects from similar classes the same way as that of objects from dissimilar classes. We propose ClasSi, a new ranking correlation coefficient which deals with class label rankings and employs a class distance function to model the similarities between the classes. We also introduce a graphical representation of ClasSi which describes how the correlation evolves throughout the ranking.

Ranking multimedia databases via relevance feedback with history and foresight support

Users of large multimedia databases which are common in scientific, commercial and entertainment applications frequently wish to explore databases sorted by their preferences. However, it is often difficult for users to explicitly express their preferences in a way that can be used directly in content-based retrieval systems. A promising approach to overcome the semantic gap between user preferences on the one hand and feature-based, quantitative similarity measures in multimedia databases on the other hand are relevance feedback systems that learn a ranking function through interacting with the user. In each iteration of a feedback loop, database objects marked as relevant are used to derive a new ranking function. To this end we propose to utilize not only the last iteration of relevance feedback but the history of all objects selected as relevant throughout the entire relevance feedback session. Using the same mathematical framework, we also examine reducing the iteration count in exploratory searches via taking the direction of movement through the database into account.

Relevance feedback for the Earth Mover's distance

Expanding on our preliminary work [1], we present a novel method to heuristically adapt the Earth Movers Distance to relevance feedback. Moreover, we detail an optimization-based method that takes feedback from the current and past Relevance Feedback iterations into account in order to improve the degree to which the Earth Movers Distance reflects the preference information given by the user. As shown by our experiments, the adaptation of the Earth Movers Distance results in a larger number of relevant objects in fewer feedback iterations compared to existing query movement techniques for the Earth Movers Distance.

Approximation techniques to enable dimensionality reduction for voronoi-based nearest neighbor search

Utilizing spatial index structures on secondary memory for nearest neighbor search in high-dimensional data spaces has been the subject of much research. With the potential to host larger indexes in main memory, applications demanding a high query throughput stand to benefit from index structures tailored for that environment. “Index once, query at very high frequency” scenarios on semi-static data require particularly fast responses while allowing for more extensive precalculations. One such precalculation consists of indexing the solution space for nearest neighbor queries as used by the approximate Voronoi cell-based method. A major deficiency of this promising approach is the lack of a way to incorporate effective dimensionality reduction techniques. We propose methods to overcome the difficulties faced for normalized data and present a second reduction step that improves response times through limiting the dimensionality of the Voronoi cell approximations. In addition, we evaluate the suitability of our approach for main memory indexing where speedup factors of up to five can be observed for real world data sets.