Balázs Hidasi

Data Mining Researcher

  • You are here: 
  • Home

Welcome!

Introduction

Hi and welcome to my personal page. My name is Balázs Hidasi and I'm a data mining researcher. Currently I'm researching recommender systems but I'm interested in everything that is data mining or machine learning. On this website you can find information about my current and past research and about me. The papers summing the results of these researches are accessible through the list of publications. If you are interested in my research or in a cooperation or just have some questions for me, feel free to contact me. Otherwise have fun browsing my homepage! :)

News

PhD

Today I successfully defended my PhD thesis on Context-aware factorization methods for implicit feedback based recommendation problems. The dissertation is based on the main bulk of the research of 2011-2014, including the initialization of matrix factorization, iTALS, iTALSx, ALS-CG/CD for speeding up ALS based methods and of course GFF. You can access the whole dissertation through the publications page.

Posted on: 28. June 2016

Deep learning

I've been working on applying deep learning for recommender systems for almost a year now. We started a research collaboration on this topic with our friends at Telefonica Research. We chose a problem, which is not really addressed by research, yet it is very important in practice: session based recommendations. Results were quite good, so we wrote a paper about it, which was accepted to ICLR 2016 and was presented at the conference the day before yesterday. We didn't stop and working on this topic eversince and have several interesting research projects in the pipeline.
I'm convinced that deep learning will be the next huge leap in recommender systems technology. I would be in fact really surprised if there won't be several papers on this topic at RecSys'16. If you are new to this topic I suggest to see the slides of my talk at the Gravity Meetup at Startup Safary Budapest. It is a very casual, informative style presentation about deep learning and how it can be included in recommender systems. I'll also give a more in depth talk at the 1st Budapest RecSys and Personalization Meetup about the session-based recommendations with recurrent neural networks. If you are interested in this topic enough to do your own research, I encourage you to submit a paper to DLRS 2016, the 1st Workshop on Deep Learning for Recommender Systems, co-organized by me. The workshop will be held in conjunction with RecSys'16 in Boston on the 15th of September, 2016.

Posted on: 06. May 2016

GFF is finally published

I'm delighted to announce that my paper on the General Factorization Framework was finally accepted and was published in the DMKD journal. I started working on GFF in late 2012 not long after I finished iTALS and iTALSx. The research started by realizing the similarities and differences between iTALS and iTALSx. The motivation behind a framework was to have a factorization algorithm for the context-aware recommendation task in which a data scientist can experiment with different preference models without having to reimplement slightly different versions of the same algorithm that are tailored to specific models. Indeed, a framework of this sort would had been very useful for my daily work as well. The work continue in early 2013 and GFF was extended beyond the original scope. After finishing it, GFF was used for a very interesting experiment in which I examined preference models beyond the traditionally used N-way and pairwise interaction models. I first tried to publish a paper on this topic in summer 2013. The main problem was that it was hard to clearly communicate the aim and usefulness of the framework AND the describe it properly AND include the experimentation within the page limits of a conference paper. The paper got mixed reviews from various conference submissions and we could always back-trace negative reviews to some kind of misunderstanding on the subject. After realizing that this was the main reason behind rejections, we moved towards journal submissions in early 2014. The first version was ok, but not that great and at this point I was already frustrated with the publication being very late. When the paper was then sent back for resubmission I left it sitting there for a longer while and focused on other work. I picked up the topic again in late 2014 and fully rewrote the paper. After two rounds of review it was finally accepted on 23th April 2015 and was published shortly after. If you are interested in the paper, head over to the publications page where you can access it.

Posted on: 09. May 2015

CaRR 2014 & visiting TUD

I will visit The Netherlands the end of this week to present the results of my newest research on modeling continuous context dimensions in the factorization framework at the CaRR 2014 workshop on 13th April in Amsterdam. I think that this research is really interesting and the results are really good (way better than what I originally anticipated). There is not a lot of work in this area which makes it even more exciting for me. Be sure to read the paper once it is available! The paper and the presentation will be uploaded here the day after the workshop. During my stay in The Netherlands I will also visit the Delft University of Technology to discuss research and possible collaborations with fellow researchers in the CrowdRec project. CrowdRec is an FP7 EU project that focuses on the developement of a new generation of recommender algorithms & systems that can seamlessly deal with stream recommendations, context-awareness while also leveraging the knowledge of their userbase through crowd engagement to provide more relevant recommendations. (For more details on the project visit their website.) Gravity R&D (the company I work for as a researcher) is one of the CrowdRec project partners.

Posted on: 09. April 2014

Meanwhile...

It has been a wile since the last update on my website. The reason for this hiatus is that the publication of my research is lagging behind my research by a great deal. A part of this is me being more interested in researching than writing publications, the other part is some of the reviewers misunderstanding the essence of my papers, thus I have to rewrite and resubmit them multiple times. I decided to temporarly resolve this by uploading draft versions of papers to arxiv.org. My papers on arxiv.org have been visible through my Google Scholar profile for a while, however I'd like to recommend two important papers here as well. The first paper goes by the name of "General factorization framework for context-aware recommendations". In this paper I present a general factorization framework that is capable of learning any linear factorization models, without restrictions on the number of dimension or the modeled interactions between them. It enables practitioners and researchers to experiment with novel models that are not used in state-of-the art factorization methods, but might be more accurate for a specific problem or for the context-aware recommendation task in general. I demonstrate how the framework can be used by comparing several different models in the framework for a context-aware recommendation problem with two predefined context dimensions. The framework with its extension is also compliant with the Multidimensional Dataspace model and allows us to use composite context as well (not just atomic context, that is usually used in these kind of algorithms). The other paper is titled "Context-aware recommendations from implicit data via scalable tensor factorization". This is an extended version of my paper on the iTALS algorithm (presented at ECML/PKDD 2012). It shows two methods to improve the scalability of the ALS learning and analyses their effect on recommendation accuracy. Additional experiments with seasonality are also included.

Posted on: 09. April 2014

CaRR 2013

I will be attending at the CaRR 2013 workshop on 5th February 2013 in Rome, where I will present an interesting side-project about context-aware similarities in the factorization framework. The full paper and slides will be uploaded here after the workshop.

Posted on: 04. February 2013

ECML/PKDD 2012

Will be attending (and presenting) at ECML/PKDD 2012, 24-28 September 2012 in Bristol.

Posted on: 17. September 2012

Updates

RecSys 2016 paper

I uploaded the draft version of my accepted RecSys 2016 paper. It is about novel parallel recurrent neural network architectures and their training, which can be used to add content (image and text) features to session-based recommendations. More details and materials (e.g. slides) will be added later..

Posted: 18. July 2016

PhD dissertation

My PhD dissertation can be accessed through the publiction page.

Posted: 28. June 2016

ICLR 2016

I presented my work on session-based recommendation with recurrent neural networks at ICLR 2016. This is one of the first works ever applying deep learning in the recommender systems domain. You can access the paper, the postes and even the algorithm of this research on the publiction page.

Posted: 06. May 2016

RecSys'15

I presented my work at the Doctoral Symposium of RecSys'15. The presentation and the accompanying paper and poster summarizes the core of my PhD research, including iTALS, iTALSx, ALS-CG/CD and GFF. I uploaded all materials to the publications section. It is a great starting point if you'd like to get an introduction into my research before reading the loger and more detailed papers. If you are interested in the details though, I advise reading the journal papers about these topics (also in the publications section).

Posted: 20. September 2015.

Another journal paper

I updated the publication list with my latest journal paper. The paper focuses on speeding up ALS learning of context-aware factorization algorithms. It introduces two approximate learning strategies and compares them to each other and to ALS. The paper was published in the prestigious KAIS journal.

Posted: 19. July 2015.

Journal papers & CV update

I updated the publication list with two journal papers.
The first one is about iTALSx, a variation of the context-aware factorization algorithm iTALS. The paper is a public version of an older tech report on the topic. It concentrates on the comparison between the two methods. The paper was published in December 2014 in the Infocommunications Journal.
The second one is about the General Factorization Framework (GFF) that I consider amongst my more important works (along with ShiftTree and iTALS). It was published in the prestigious DMKD journal a few days ago. It took a while to publish this topic (see post in the news for details), but I think it was certainly worth the hassle.
I also updated my CV page.

Posted: 09. May 2015.

Presentations from The Netherlands & paper

I updated the publication list with my paper for CaRR2014 (& presentation). I already wrote a little bit about this paper here in the news section and I still recommend reading it. Beside CaRR, I also gave a talk at the Delft University of Technology about my research in general. The slides of this presentation are available through my SlideShare profile.

Posted: 13. April 2014.

My first journal paper

List of publications was updated with my recently published paper titled "Initializing Matrix Factorization Methods on Implicit Feedback Databases". It was published in J.UCS (Journal of Universal Computer Science) and is a significantly extended version of my CaRR 2012 paper. The full text is freely accessible on the J.UCS website. (I put a direct link to it on the publication page.)

Posted: 14. October 2013.

Profiles

Links to my various profiles (including LinkedIn, SlideShare, etc) are now provided on the "Contact" page. Additionally there were some chages in the design of the site.

Posted: 06. May 2013.

Double update

  • Resources about context-aware similarities uploaded (CaRR 2013 paper and slideshow). You can access them through my publication list or the Research/Recommender Systems menu.
  • All slides now are available on SlideShare as well as the original PPT for your convenience. Please note that SlideShare doesn't support animations therefore animated slides were splitted into multiple slides.

Posted on: 11. February 2013.

iTALS resources

Resources about iTALS - presented at ECML/PKKD - are now up (draft version of paper, slides, poster).

Posted on: 29. September 2012.

Website is up

It took a long time, but the first version of this page is finally up & working. The description of the research areas will be expanded soon.

Posted on: 17. September 2012