Case: Easy access to processing power for computational recommendation models(Publicatie)

Together with a group of master's students, university lecturer and researcher Flavius Frasincar from Erasmus University Rotterdam (EUR) is researching how recommendation systems for online financial news can be improved. Systems of this type automatically recommend news items to readers based on their reading profiles.

download
24 AUG 2016

Complex models

“Earlier models were pretty simple,” explains Flavius Frasincar. "They only looked at the words present in a news item and made recommendations to the reader on that basis. Now we are also including the meaning behind words, i.e. homonyms, synonyms and associations – all possible relationships between words." He gives the example of the ECB and Draghi: names that don't feature in a dictionary but which nevertheless belong to each other.
Erasmus Universiteit SURFsara training RCCS
Photography: Arie Kers, Erasmus University Rotterdam

Computations not feasible on a PC

"All these relationships between words mean that we have a great many parameters for our computations. We also include various threshold values which determine whether or not the system recommends a news item." This necessitates a huge number of computations. Attempts to perform these computations on a PC failed. "After a couple of weeks the computer still hadn't finished. One of my students tried this but it quickly became clear that we needed more processing power.” It wasn't so much the size of the dataset, which at around 100 news items wasn't particularly big. Computing more than thirty parameters for each threshold value in combination with the processing of natural language made this task computationally difficult, explains Frasincar.

Fast access to processing power

SURFsara's Lisa compute cluster soon came to mind. "I was already using the Lisa cluster through an NWO project. But halfway through the year, our research group had already used up 98% of the allotted computing hours. So it's great that the EUR is now also purchasing computing time from SURFsara. If I need extra computing hours this can be arranged quickly and easily. All I have to do is email Marlon Domingus in EUR's Research Support Office.” Domingus coordinates the setup of the High Performance Computing (HPC) service. “The director of the University Library, Matthijs van Otegem, has enabled access to the High Performance Computing service,” says Domingus. “This year he is financing a pilot project to identify the demand for high-performance computing and to find out whether we can offer this service as standard to EUR's researchers. That way we'll know how many computing hours are required and which HPC systems will be used for computing. This first year has clearly been a learning curve. Initially we only offered the Lisa cluster, but now we also offer the national supercomputer Cartesius and recently the HPC Cloud as well.”

Getting started

It was also Domingus who, in conjunction with SURFsara, organised an introductory course in Cluster Computing for EUR's researchers. “This introduction was essential to enable my students to get started with the Lisa compute cluster,” says Frasincar. "Obviously, you need a certain amount of knowledge before you can start computing with Lisa. My students were familiar with Java, but they didn't know anything about Linux, the operating system that the Lisa compute cluster runs on. We also took full advantage of the support offered by SURFsara during the research project. That's when we noticed that we had quite a few specific questions. For example, we needed technical expertise to configure a new version of a Java virtual machine on the Lisa cluster. SURFsara gave us a great deal of help with this, and everything was sorted really quickly."

RCCS contract with SURFsara

Following a positive review of the pilot year, the University Library and the University Support Centre (USC) IT have renewed the contract with SURFsara. This means that SURFsara's HPC services will continue to be available to EUR's researchers over the coming year.

More information

Result of:Support4research
Number of times shown:
1046
Latest modifications 27 Sep 2018