[BI Announcement] ICAART 2019: Reza Refaei Afshar won ”Best Student Paper Award”

Reza Refaei Afshar, PhD student at the Information Systems group, has received the Best Student Paper Award for his work presented at the 11th International Conference on Agents and Artificial Intelligence held on 19 – 21 February, in Prague, Czech Republic. The paper titled “A Reinforcement Learning Method to Select Ad Networks in Waterfall Strategy” was co-authored by his supervisors Yingqian Zhang, Murat Firat and Uzay Kaymak.


Nowadays, one of the most important sources of income for publishers who own websites is through online advertising (online ADs). For online publishers, it is difficult to design good strategies to manage their online AD auction systems due to highly dynamic real-time bidding environment.  This paper proposes a machine learning based decision support system for publishers, which is built from historical AD auction data. The proposed method demonstrates its effectiveness in terms of the increased expected revenue for publishers.

BPI cluster meeting 30Oct’18

This week, October 30, 12:30-13:30h, in Pav. K16, we have our cluster meeting. Paulo will give a talk on “Remaining Useful Lifetime (RUL) prediction across different domains”. Below you can find the abstract of the talk.

Remaining Useful Lifetime (RUL) prediction across different domains:

X-ray tubes are critical components in scanners used for radiography procedures. Tubes are composed of tungsten filaments which are heated up to produce x-rays images. However, such filaments are subject to wear-out given intensive usage. Such behaviour can lead to an early failure of the filaments causing disruptions and incurring high maintenance costs. Currently, there are two types of scanner generators in operation; the newest generator can log enough sensor information for the construction of a Physics Model capable of estimating the RUL. However, older versions do not store crucial sensor data to enable a prediction. Inspired by this use case, we evaluate a transfer learning method to predict the RUL for an equipment when there is a shift in the input distribution among different data sources. We test the method on a benchmark data with different failure modes and unlogged RUL information.


BPI cluster meeting 16Oct’18

This week  Daniel Kapitan from Mediquest company will give a talk during our BPI cluster meeting. Timing is between 12:30-13:15 on 16th October 2018 in  room Pav. K.16.

The topic of the talk will be related to Healthcare and the title is

Project Nightingale: Machine learning as a catalyst for outcome-oriented care

Project Nightingale applies machine learning within the context of value-based healthcare. Given the need for better insight into outcomes, and the potential of predictive analytics for clinical decision support, results from pilot projects are presented, where:

* Compounded outcome measures relevant for shared-decision making are defined, using existing data dictionairies (ICHOM)

* Imbalanced learning is applied to identify high-risk patients prior to an intervention

* Results of machine learning are related to existing epidemiological research

Ultimately, the aim is to promote the use and understanding of machine learning in outcome-oriented healthcare.


BPI cluster meeting 18Sept’18

This week  Emmanuel Rios Velazquez from IMEC Netherlands, will give a talk during our BPI cluster meeting. Timing is as usual between 12:30-13:30 on 18th September in  room Pav. K.16. You can find further information about Emmanuel’s work below.

Digital health: Tracking mental stress and mood using wearable sensors and machine learning.

Through wearable technology and digital footprints on our mobile phones, social media, etc., we can extract valuable insights on our lifestyle and well-being. imec is developing new tools to encourage behavior change towards a healthier lifestyle, that may lead to new therapeutic tools for patients with mental health problems such as depression or eating disorders.

Through habit monitoring, physiological monitoring, and personalized, intelligent algorithms we aim to identify triggers of unhealthy behavior, increase awareness and contribute to preventive health.

This talk will touch on data-driven machine-learning enabled, health and virtual reality applications to detect cognition, mood and behavior using wearable data.



Emmanuel is a data scientist engaged on the identification of relevant patterns on physiology and brain activity to assess cognition, mood and behavior through wearable technology, at Imec’s Connected Health Solutions team.

Emmanuel received a PhD from the Maastricht University, on the use of heterogeneous patient data and computational imaging (radiomics) for decision-support systems in radiation oncology.

This pioneering work led to diverse multi-centric, international scientific collaborations. This was followed by a post-doctoral degree at the Computational Imaging and Bioinformatics lab at the Dana-Farber Cancer Institute-Harvard Medical School, investigating the link between cancer imaging phenotypes and tumor biology for precision medicine.

Besides crunching data, he likes reading (contemporary novels), playing drums and whenever possible going to the sea.

BPI cluster meeting 4July’18

This week Guillaume Crognier will give a talk during our BPI cluster meeting. Timing is as usual  between 12:30-13:30 in the room Pav. K.16. Guillaume will present the outcome of his work since he joined our group.


Constructing decision trees by using Column Generation with restricted parameters


Almost every machine learning algorithm to generate decision trees is greedy, and may be far from the « optimal » decision tree. Research papers have already tried to model this problem as a MILP (mixed integer linear program), but most of them are too slow to be used in practice or cannot deal with big datasets. The purpose of this work is to show that such optimal algorithms can be greatly improved (considering the quality of the tree as well as the computational time) by using columns generation.

You are warmly invited to Guillaume’s talk.

BPI cluster meeting 20June’18

This week we will have a guest presenter,  dr. S. Faghih Roohi from OPAC group of our faculty. The place of the presentation is Pav.K1.6 and time is between 12:30-13:30.


A group decision making approach for risk ranking and lane selection in distribution of pharmaceutical products


This study aims to provide a group decision making framework based on the prioritized risks for selecting shipment lanes of pharmaceutical products. The risks involved in the decision making are identified and categorized by referring into the conventional failure modes and effect analysis (FMEA) tables. Using categorized risks, a new shortened FMEA table is proposed for evaluation by a group of experts in pharmaceutical distribution and logistics industry. The evaluations by experts are primarily in linguistic terms which are further converted to intuitionistic fuzzy numbers (IFNs) for aggregation operations. By using an intuitionistic fuzzy hybrid TOPSIS (Technique for Order Preference by Similarity to Ideal Solution) approach, the risks in shipment processes for every lane are scored and prioritized. The relative closeness coefficients of risk categories resulted from intuitionistic fuzzy hybrid TOPSIS are used for lane comparison and selection. Lane selection is performed in multiple rounds in a way that in each round, the higher-scored lanes are selected based on the lower-ranked (higher-priority) risks. The proposed approach provides an opportunity for all managers and decision makers to evaluate risks and to keep/establish current/new lanes. Finally, a case study of lane selection on air cargo distribution of pharmaceutical products is presented to demonstrate the potential applications of the proposed approach.

You are warmly invited.

BPI cluster meeting 30May’18

This week we will have a guest presenter, Gabriele Modena from ImproveDigital company. The place of the presentation is Pav.K1.6 and time is between 12:30-13:30.


Machine Learning Methods in Adtech

Short Abstract:

Machine Learning systems are used in adtech to drive decision making, revenue, and personalise the user’s online experience. ML is used to answer questions such as is the visitor a human or a bot? What creative should be displayed in order to maximise the probability of a click? What is the optimal reserve price for an impression? How many clicks will a new ad-placement system get?

Typically we need to answer these questions several tens of thousands of times per second, under soft real-time constraints.

In this presentation we’ll give an overview of use cases in the industry, and how machine learning is used at scale in the Improve Digital platform.

 About the company:

Improve Digital has the All-in-One Advertising Platform for Publishers, Content Providers and Broadcasters. Improve Digital annouces its mission as building smart, efficient, and responsible digital businesses for its enterprise customers. It creates the technology that makes advertising marketplaces possible.

You are warmly invited.

BPI cluster meeting 23May’18

This week we will have a guest presenter, dr. Kalliopi Zervanou, is currently a lecturer in Information and Computing Sciences in Utrecht University. The place of the presentation is Pav.K1.6 and time is between 12:30-13:30.


Linking multi-disciplinary data sources in the Time Capsule system



Time Capsule is a historical research system for botanical remedies from the New World in the early modern period (17-18th century): the historical evolution of economic importance, ethical attitudes, scientific interests, trade and knowledge circulation.


Historical data is scattered across collections developed for various domains and purposes. Its amount and complexity raises the need for a presentation allowing exploration and detailed inspection. Finally, the problem of information validation and sharing must be addressed. In Time Capsule we have i) integrated and linked multidisciplinary data sources and ii) developed an online research platform that supports data access, presentation, validation and sharing.


In our approach, data source integration entails concept mapping, not only across disciplines, but also in time. Thus, it calls for support for the scientific evolution from the 16th century onwards in re-classifying and re-defining concepts. Additionally, it entails dealing with phenomena of historical term variation and ambiguity which gradually give way to spelling standardisation and current nomenclature conventions in e.g. botany and biology. Furthermore, it requires addressing under-specificity and ambiguity of information found in historical sources while maintaining associations with potentially related concepts and context. Most importantly, it requires providing references for information provenance tracing and validation.

You are warmly invited.





BPI cluster meeting 2May’18

Our cluster meeting on 2nd May will be in Pav. K.16 between 12:30-13:30. This week we will have a guest lecturer Maurits Kaptein

The title and the abstract of Maurits Kaptein’s lecture are as follows:


Personalization and bandits with applications in health.



In this talk prof. Maurits Kaptein will talk about his work in the computational personalization lab at JADS (Den Bosch). Starting from a (contextual) multi-armed bandit formalization of treatment personalization Maurits will present his work on developing novel multi-armed bandit policies (e.g., bootstrapped thompson sampling), on software development to evaluate multi-armed bandit policies in the field (https://github.com/Nth-iteration-labs/streamingbandit), and on the use of this framework for personalization in (e)Health and online marketing. The talk will provide a broad overview of the work carried out by Maurits and his PhD students over the last five years.


Short Bio of Maurits Kaptein:

He received his Ph.D. with honors form the Eindhoven University of Technology, Eindhoven, the Netherlands. Next, he worked as a postdoctoral researcher at the Aalto school of Economics, Aalto, Finland. Afterwards he worked for 2 years as an assistant professor of Statistics and Research Methods at the University of Tilburg. He has previously (during his Ph.D. work) worked as a research scientist at Philips Research, Eindhoven, the Netherlands and as a distinguished visiting scholar at the CHIMe lab of Stanford University, Stanford CA, USA. He has also worked as an assistant professor in Artificial Intelligence (AI) at the Radboud University Nijmegen where he was the track leader of a master track called “Web and Language”.

BPI cluster meeting 25Apr’18

Our BPI cluster meeting on 25th  April will be a joint meeting with Data Science Center Eindhoven (DSC/e).

The DCS/e lecture will be given by Marie-Jeanne Lesot. She is an associate professor in the department of Computer Science Lab of Paris 6 (LIP6) and a member of the Learning and Fuzzy Intelligent systems (LFI) group.

Where: Kennispoort (Grote Zaal), J.F. Kennedylaan 2, Eindhoven.

When: 25th  April, as usual, between 12:30-13:30. (doors open at 12:00).

The title and the abstract of Marie-Jeanne Lesot’s lecture is as follows:


Extracting knowledge in linguistic form


Machine learning can be seen as aiming to allow users to understand the huge quantities of data they are faced with. One way to facilitate interpretation of the results consists in presenting them in natural language, offering linguistic expressions which may be easier to understand. The choice of such result formulation then has an impact on the machine learning techniques to be applied to the data. This talk will present three tasks in this framework, considering different types of data.

The first task aims at extracting gradual itemsets from numerical data, as well as contextual variants thereof, linguistically expressing information about the feature covariations, as illustrated by the example “the higher the speed, the greater the danger”. A second task aims at summarising temporal series, in particular their periodicity, using the specific quantifier “regularly”. In both cases, the question is to precisely define the associated semantics and to define efficient extraction algorithms.  A third task investigates the measure of the relevance of the linguistic terms used to express the summaries, both with respect to the data structure, in case of linguistic variables, and with respect to the cognitive interpretation, in case of approximate numerical expressions.

Short bio:

Marie-Jeanne Lesot obtained her PhD in 2005 from the University Pierre et Marie Curie in Paris. Since 2006 she has been an associate professor in the department of Computer Science Lab of Paris 6 (LIP6) and a member of the Learning and Fuzzy Intelligent systems (LFI) group. Her research interests focus on fuzzy machine learning with an objective of data interpretation and semantics integration and, in particular, to model and manage subjective information; they include similarity measures, fuzzy clustering, linguistic summaries, affective computing and information scoring.