Publications by type - Jeroen Janssens

Publications

Journal Papers

2010

Willems, C.M.E., W.R., van H., G.K.D., de V., Janssens, J.H.M. and Malaisé, V. An integrated approach for visual analysis of a multi-Source moving objects knowledge base. In International Journal of Geographical Information Science, 2010. (show abstract)
We present an integrated and multi-disciplinary approach for analyzing behavior of moving objects. The results originate from ongoing research of four different partners from the Dutch Poseidon project (Embedded Systems Institute (2007)), which aims to develop new methods for Maritime Safety and Security (MSS) systems to monitor vessel traffic in coastal areas. Our architecture enables an operator to visually test hypotheses about vessels with time-dependent sensor data and on-demand external knowledge. The system includes the following components: abstraction and simulation of trajectory sensor data, fusion of multiple heterogeneous data sources, reasoning, and visual analysis of the combined data sources. We start by extracting segments of consistent movement from simulated or real-world trajectory data, which we store as instances of the Simple Event Model (SEM), an event ontology represented in the Resource Description Framework (RDF). Next, we add data from the web about vessels and geography to enrich the sensor data. This additional information is integrated with the representation of the vessels (actors) and places in SEM. The enriched trajectory data is stored in a knowledge base, which can be further annotated by reasoning and is queried by a visual analytics tool to search for spatio-temporal patterns. Although our approach is dedicated to MSS systems, we expect it to be useful in other domains.

Conference Papers

2010

Janssens, J.H.M., Hiemstra, H. and Postma, E.O. Creating artificial vessel trajectories with Presto. In Proceedings of the 22nd Benelux Conference on Artificial Intelligence (BNAIC 2010), Luxembourg, October 2010. (show abstract)
The automatic detection of anomalies in the maritime domain requires representative anomalous instances. We developed Presto, an application that enables maritime-domain experts to create artificial anomalous vessel trajectories that may be characteristic of traffic violations, illegal fishing activities, drug smuggling, or piracy. When merged with existing real-world data, the artificial trajectories make possible the usage and evaluation of machine learning algorithms for the automatic detection of anomalies.
Willems, C.M.E., van Hage, W.R., de Vries, G.K.D., Janssens, J.H.M. and Malaisé, V. An integrated approach for visual analysis of a multi-source moving objects knowledge base. In AGILE workshop on Geospatial Visual Analytics: Focus on Time, Guimarães, Portugal, May 2010. (show abstract)
We present an integrated and multi-disciplinary approach for analyzing behavior of moving objects. The results are ongoing research of four different partners in the Dutch Poseidon project [3] where we aim for new developments for Maritime Safety and Security (MSS) systems to monitor vessels. We focus on the following requirements for an MSS system: abstraction of large amounts of trajectory sensor data, fusion of multiple heterogeneous data sources, and visual analysis of the combined data sources. We start by extracting segments of consistent movement from trajectory data, which we store as instances of the Simple Event Model (SEM), an event ontology represented in the Resource Description Framework (RDF). Then we add data from the web about vessels to enrich the sensor data. This additional information is integrated with the representation of the vessels (actors) in SEM. The enriched trajectory data is stored in a knowledge base, which is queried by a visual analytics tool to search for spatio-temporal patterns. Although our approach is dedicated to MSS systems, we expect it to be useful in other domains.

2009

Janssens, J.H.M., Flesch, I. and Postma, E.O. Outlier detection with one-class classifiers from ML and KDD. In Proceedings of the Eighth International Conference on Machine Learning and Applications, Miami, FL, USA, December 2009. (show abstract)
The problem of outlier detection is well studied in the fields of Machine Learning (ML) and Knowledge Discovery in Databases (KDD).
Both fields have their own methods and evaluation procedures. In ML, Support Vector Machines and Parzen Windows are well-known methods that can be used for outlier detection. In KDD, the heuristic local-density estimation methods LOF and LOCI are generally considered to be superior outlier-detection methods. Hitherto, the performances of these ML and KDD methods have not been compared.
This paper formalizes LOF and LOCI in the ML framework of one-class classification and performs a comparative evaluation of the ML and KDD outlier-detection methods on real-world datasets.
Experimental results show that LOF and SVDD are the two best-performing methods. It is concluded that both fields offer outlier-detection methods that are competitive in performance and that bridging the gap between both fields may facilitate the development of outlier-detection methods.
Janssens, J.H.M. and Postma, E.O. One-class classification with LOF and LOCI: An empirical comparison. In Proceedings of the 18th Annual Belgian-Dutch Conference on Machine Learning, pages 56-64, Tilburg, The Netherlands, May 2009. (show abstract)
LOF and LOCI are two widely used density-based outlier-detection methods. Generally, LOCI is assumed to be superior to LOF, because LOCI constitutes a multi-granular method. A review of the literature reveals that this assumption is not based on quantitative comparative evaluation of both methods.
In this paper we investigate outlier detection with LOF and LOCI within the framework of one-class classification. This framework allows us to perform an empirical comparison using the AUC performance measure.
Our experimental results show that LOCI does not outperform LOF. We discuss possible reasons for the results obtained and argue that the multi-granularity of LOCI in some cases may hamper rather than help the detection of outliers.
It is concluded that LOCI does not outperform LOF and that the choice for either method depends on the nature of the task at hand. Future work will address the evaluation of both methods with existing one-class classifiers.

Masters Thesis

2008

Janssens, J.H.M. Collaborative image ranking: Bridging the semantic gap with human computation. Masters thesis, Maastricht University, January 2008. (show abstract)
Most image search engines use a ranking algorithm which assigns to each image the relevance value of the associated webpage. The resulting image ranking is of limited value because images, unlike webpages, lack a semantically rich underlying hyperlink structure. A sensible alternative would be to rank the images according to their content. In the thesis, we define the semantic ranking problem (SRP) as the problem to obtain a ranking of images of a certain class, based on a single appropriate semantic attribute. Current computer-science techniques cannot extract the appropriate semantic attributes of an image. This well-known problem is called the semantic gap. It may be bridged by using human computation. Human computation is a technique where certain steps of a computational process are outsourced to humans.
The thesis focusses on the following problem statement: To what extent is human computation a suitable approach to solve the SRP? To investigate the problem statement, we present the CollaboRank method. CollaboRank enables humans to rank a set of images collaboratively, by means of four steps: (1) divide the SRP into appropriate tasks, (2) let humans perform the tasks, (3) aggregate the answers, and (4) produce a global ranking. For the steps (1) and (3), two task-formulation algorithms and two ranking-aggregation algorithms were developed. For step (4) we adopted the Greedy-Order algorithm of Cohen, Schapire, and Singer (1999). The performance of CollaboRank was evaluated using a simulation environment with controlled parameters, where the humans are simulated by agents. Furthermore, a validation experiment was conducted where 102 people collaboratively ranked three sets of images.
Our experiments showed that CollaboRank is scalable in terms of image-set size, and that it can handle objective as well as subjective semantic attributes. The experiments also indicated that the performance of CollaboRank can be improved using intelligent task-formulation and ranking- aggregation algorithms. They further demonstrated how to compute the appropriate task size in such a way that the number of human computations is optimised. From the obtained results, we may conclude that human computation is a suitable approach to solve the SRP.