You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@mahout.apache.org by Felipe Martins <fe...@yahoo.com.br> on 2011/03/21 14:46:55 UTC
Proposal - Idealize Recommendation Framework
Hi everybody,
First and foremost, thank you very much for reading this e-mail.
My name is Felipe and I’m master student at Federal University of Ouro Preto
(www.ufop.br) in the field of Information Retrieval. There we have a lab called
Idealize where several projects in that field are hosted. One of these projects
is the reason for this e-mail.
We have developed a project in partnership with upLexis Technology
(http://www.uplexis.com.br/) in the area of recommender systems. Our intention
has been to develop recommendation applications following different approaches,
for which we decided first to put some effort into creating a framework for
general-purpose recommender systems. This framework was named Idealize
Recommendation Framework (IRF) after lab’s name. IRF has benefited a lot from
Mahout, mainly Taste, which has provided us a ground from where the
collaborative filtering features of our systems have been developed. We have
reused practically all recommendation methods provided by Taste and our
applications have been achieving exciting results. So here goes a “thank you
very much” for all of you involved in the Mahout Project.
However, we have been facing several demands that ask for other approaches to be
used, and unfortunately, collaborative filtering is not suitable for all of
them. For instance, we have demands for recommending news, people based on their
physical traits, products by their metadata and so on. Nevertheless, new
recommendation approaches needed to be considered.
As already stated, departing from Mahout(Taste), we have devised IRF for
general-purpose recommender systems, that means, IRF is intended to support
recommendation applications based in any approach. IRF is both a conceptual and
a software framework, and we have achieved remarkable reduced time-to-market for
new applications developed on top of IRF as well as very good performance in
production environment due to the following:
. IRF allows the notion of a process so that the required steps to create new
applications are already well established;
. IRF devises a production model which improves overall performance in a very
exciting way;
. IRF architecture allows a great deal of reuse;
Also, we are currently developing chart and report components to integrate with
IRF, as well as tools for automatically deploy, an Eclipse plug-in to
automatically create the hotspots for each sector (see attached file), and, of
course, new methods and applications.
We have already developed several applications based on different approaches on
top of IRF. We have content-based applications, usage-based applications and
hybrid ones.
After being successful in developing and deploying recommender systems, we
decided it would be great and fair to make IRF available so that the community
could benefit from it, exactly the same way we benefited from open-source
software to develop IRF.
As we make extensive use of Apache software (Mahout because of Taste and machine
learning algorithms, Lucene, Hadoop and HBase to assist recommendation methods,
Nutch and Tika to collect metadata from Web, Tomcat as the container holding the
recommendation applications, Ant as a deploy tool, and even Struts and Tiles
when we develop the front-end), nothing more natural than considering Apache the
first place to make IRF available.
I’ve already exchanged some e-mails with Sean Owen, who kindly elucidated me
about several issues regarding Apache projects and suggested me to send this
e-mail in order to collect community’s impressions and insights.
As IRF builds on Taste, we, at Idealize lab, guess it could maybe be contained
within Mahout, as a Taste’s “brother”, for instance.
We wrote a paper about IRF which was accepted by the 14th International ACM
SIGSOFT Symposium on Component Based Software Engineering, which will take place
in Boulder, Colorado, next June. The paper is attached so that you can have
deeper access to the technical issues.
Finally, we have a great team here at Idealize lab. All the members of our team
are very excited about the possibilities with IRF and would become a very strong
initial workforce.
I’d like to thank you all for reading this e-mail and say that we, at Idealize
lab, are very hopeful about IRF and waiting to hear from you.
Best regards,
Felipe.