Predictive-Analytics.be

Bart Hamers, Ph.D.

 The greatest difficulty in the world is not for people to accept new ideas, but to make them forget old ideas."
John Maynard Keynes

My name is Bart Hamers. I am a Professional Analyst with both strong academic and business experience.  My main sphere of interests is Information Management for Predictive Analytics.  I am interested how companies use  their Information and Data in the most efficient and way. How should information be captured, stored, treated, used  and communicated with as final goal efficient decision support.

Information Management for Predictive Analytics  overlaps with different  sub domains:  Data Governance, Business Intelligence, Data Mining,  and Statistical and Mathematical Modeling. I had the opportunity to apply these techniques in Banking, Marketing and Telco. 

Recently I am interested in the data governance for quantitative business solutions  in terms of both IT, mathematical and project mode. How should one structure a quantitative decision support team within an operational business environment. This involves typically 3 phases:

All of them have to align in order to have highest efficiency of your project. But also the project management of quantitative modeling projects is al challenging topic. Less is known about this topic since data mining projects typically live between IT and R&D.

Since April 2006:  I am working for the Risk Management Department of the Dexia Group as a Senior Quantitative Analyst.

My job focus lies in in the domain of Credit and Operational Risk.  My main tasks are

March 2005 -Januari 2006:  IKAN Consulting, Marketing Data Mining Consultancy  in Telco,

March 2004 - March 2005: Vadis Consulting (Part of the WegenerDM direct marketing group), Marketing Data Mining Consultancy  in Telco and Banking,

August 1999-March 2004: Teaching and Research Assistant at the Department of Electrical Engineering, KUL (ESAT/SCD-SISTA)

Kernel Models for Large Scale Applications

As a result of the ever-growing influence of IT in research and companies terabytes of data are handled and stored daily. Therefore the interest in using this source of information is increasing steadily. It is there that data mining and machine learning gives us a hand. One recently developed set of tools in data mining are kernel models. These models excel in a variety of problem situations like classification, regression, time-series prediction problems with respect to the generalization performance tested on unseen data. The disadvantage of these models is that the computational demands for training them scale quadratically with the size of the training set. In this work we will show how this scalability problem for kernel models can be overcome by making use of a combination of methods from numerical and learning theory origin.

The proposed solutions, which have to be taken into account when one wants to use kernel models on large scale applications, will be presented on the basis of five pillars. These five pillars are: the model choice, the numerical procedures, the choice of the kernel, low rank approximations and the use of ensemble models. First we will show how models using a quadratic loss function will have a training procedure that consists of a linear system. We will show how iterative methods and low rank approximations can reduce the computational and memory complexity for solving this linear system. Also we will show that the kernel itself plays an important role in both the learning performance and the computational and memory complexity of the algorithms.

As a last solution we will propose the use of ensemble models for training kernel models on large data sets. Instead of training one model on a training set, a whole set of models is trained on subsets of the original training set. In this way the scalability problem can be avoided. In addition we will introduce a new method of coupled learning. In this methodology the members of the ensemble will share the knowledge during training. This leads to a new way of transductive learning.

Hamers B. "Kernel Models for Large Scale Applications'', PhD , Katholieke Universiteit Leuven, Belgium,2004. 

(Download from ESAT ftp-server pdf / ps / ps.zip)

The results of this research have contributed in the development of the Matlab LS_SVMlab package. To download this free software package visit the webpage.

Hamers B., Suykens J.A.K., Leemans V. De Moor B., ``Ensemble Learning of Coupled Parameterized Kernel Models'',in Proceedings ICANN/ICONIP 2003, Istanbul, Turkey.

Hamers B., Suykens J.A.K., De Moor B., ``Coupled Transductive Ensemble Learning of Kernel Models'', Internal Report 03-172, ESAT-SISTA, K.U.Leuven (Leuven, Belgium), 2003.

Hamers B., Suykens J.A.K., De Moor B., ``A comparison of iterative methods for least squares support vector machine classifiers'', Internal Report 01-110, ESAT-SISTA, K.U.Leuven (Leuven, Belgium), 2001.

Pelckmans K., Suykens J.A.K., Van Gestel T., De Brabanter J., Lukas L., Hamers B., DE Moor B., Vandewalle J., `` L S-SVMlab : a Matlab/C toolbox for Least Squares Support Vector Machines'', Internal Report 02-44, ESAT-SISTA, K.U.Leuven (Leuven, Belgium), 2002.

Gerd Castermans, David Martens, Tony Van Gestel, Bart Hamers, Bart Baesens, An Overview and Framework for PD Backtesting and Benchmarking, CRC conference,  2007.   (Accepted for Publication Journal of the Operational Research Society)

Karel Dejaeger, Jessica Ruelens, Datakwaliteit in Risk Management, 2008-2009.

Afef Ben Khaled, Statistical Analysis of Operational Risk Data, VIE, 2008-2009.

Bart Hamers

Konongin Astridlaan 42, 3010 Leuven, Belgium

mobile: +32 479 90 06 67

Email:  follow the link

You can also find me by

Informative Sites about AI, Neural Networks, Machine Learning and philosophy

Hobbies and Fun:

Also visit the site of my beautiful children Siebe and Cléo

I'm also a big fan of: