Skip to content

Workshop on Academic-Industrial Collaborations for Recommender Systems

May 30, 2013

Workshop Date: 10 July 2013

Submission Date: 1st July, 2013

Location: Headrooms, 1-2 St John’s Path, London, EC1M 4DD

What is this Workshop?

Recommender systems are fast becoming as standard a tool as search engines, helping users to discover content that interests them with very little effort.  Their commercial popularity has made them as hot a topic in industry as they are in academia.  With this shared interest, conferences like ACM’s Recommender Systems often attract as many academics as industry practitioners.  This workshop aims to bring local academics and industry practitioners together to share experiences and make new contacts.

This workshop provides a forum for researchers and developers alike to showcase their recommender system work and explore possibilities for new collaborations.  We’re interested in fostering new links between academia and industry, providing academics with real-world environments in which to solve problems and industrial partners with access to some of the local area’s top academic researchers in the field.  We’re also interested in hearing from researchers and developers who are currently or were previously involved in such collaborations, learning from what went well and not so well in their experiences.

Call for Presentations/Demonstrations

Researchers and developers from academia and industry are invited to give presentations in our 1-day workshop.  We expect each presenter will have a total of 20 minutes to present and/or demonstrate their recommender system work which will be followed by 10 minutes of discussion.  Topics of interest include, but are not limited to:

  • Academic’s experiences of working with industry
  • Industry practitioner’s experiences of working with academics
  • Experiences of working with recommender system software libraries
  • Experiences of testing recommenders offline
  • Experiences of testing recommenders online
  • Experiences of sharing data between academia and industry

If you would like to present, please contact kris.jack@mendeley.com with a short description of your presentation in no more than 300 words.  Also, please let us know if you are already involved in any academic-industry partnerships.  If so, what were the main challenges that you had in working together and how, if possible, did you overcome them?  If not, are you seeking some partners?  What would you like to get from the partnerships?  This will help us to identify useful topics for our round table discussion.

Important Dates

Presentation Abstract Deadline: 1st July, 2013

Notification of Acceptance Deadline: 3rd July, 2013

Workshop Date: 10th July, 2013

Please let us know (kris.jack@mendeley.com) if you will attend the workshop whether you intend to give a presentation or not.

On receiving your abstract, we will get back to you within days as to whether we think that you’re presentation would be interesting for the workshop audience.  So if you submit before 1st July, you can expect to receive a reply sooner than the notification deadline too.

Who is expected to attend?

Researchers and developers who work with recommender system technologies.  We invite researchers who are working on novel solutions to some of the most challenging problems faced in recommender system research and industry practitioners who grapple with the challenges of developing recommender systems into products for real people to use.  By bringing researchers and industry practitioners together for a day we hope that everyone will benefit from making new contacts in an open and friendly environment.  We particularly encourage both researchers and industrial partners who have had previous experiences of academia-industry collaborations.  As it is a one day event, we expect the audience to be quite local, made up mainly of participants from London and the UK although all are welcome.

Workshop Programme

9:00 Breakfast and coffee

9:20 Open workshop and welcome

9:30 Jagadeesh Gorla, A Bi-directional Unified Model

10:00 Break

10:20 Nikos Manouselis & Christoph Trattner, Je t’aime… moi non plus: reporting on the opportunities, expectations and challenges of a real academic-industrial collaboration

10:50 Heimo Gursch, Thoughts on Access Control in Enterprise Recommender Systems

11:20 Break

11:50 Maciej Dabrowski, Towards near real-time social recommendations in an enterprise

12:20 Mark Levy, Item Similarity Revisited

12:50 Lunch

14:05 Benjamin Habegger, 109Lab: feedback from a start-up experience in collaboration with academia

14:35 Thomas Stone, Venture Rounds, NDAs and Toolkits – experiences in applying recommender systems to venture finance

15:05 Break

15:25 Round table discussion on academic-industry collaborations

16:25 Break

16:45 Exploring possible collaborations

17:30 Drinks for all at Giant Robot!

Note: Presentation slots have 20 minutes for presentations followed by 10 minutes of discussion.

Accepted Presentation Abstracts

Maciej Dabrowski, Towards near real-time social recommendations in an enterprise

http://advansse.deri.ie/

This work proposes a new perspective on near real-time social recommendations in enterprise social platforms based on the challenges we identified in a comprehensive survey of existing approaches as well as the enterprise social platforms. We explore how Linked Data techniques associated primarily with the Semantic Web can be applied in a Corporate Enterprise context to provide near real-time social recommendations. We draw from research and experimental experience along with practical experiences based on the Webex Social platform, which is modeled on three elements: people, communities, and information. We evaluate a number of potential recommendation and notification techniques before proposing a novel method for generating recommendations based on the concept of Spreading Activation in an enterprise social graph. Our approach considers multiple information sources and is based on an experimental demonstrator using data from DBPedia and other sources.  The proposed method allows for dynamic inclusion of additional information sources with very low effort, significantly improving recommender flexibility. Furthermore we define a system and architecture capable of making recommendations in near real-time whereas the majority of existing recommendation systems focus on a pre-compute preparation of results. This is critical in enterprise environments where informational events but more importantly their context and meaning is time critical. The architecture that we propose may be deployed in a standalone mode or as an integral part of an enterprise social software platform. We highlight how our approach improves the management of knowledge and communication within organization through more effective exploitation of information available in the corporate social web.

In addition to the contribution of this architecture and system which contains a novel approach for enterprise social recommendations we also propose the extension of certain IETF and W3C standards, namely the eXtensible Messaging and Presence Protocol (XMPP) and Semantic Processing And RDF Query Language (SPARQL) respectively.

Jagadeesh Gorla, A Bi-directional Unified Model

The underlying problem of recommender systems is to present a user with a set of relevant items (products, ads, people, research papers, etc.). The simplest way to look at it as the problem of estimating the probability of relevance between the user-item pairs and ranking them based on the probability. Even though it seems simple, it poses a difficult challenge on how to utilise  the available information about the users, items, and the interactions between various user-item pairs in estimating the relevance. A simple example would be, in collaborative filtering we only have information on the user-item interactions (clicks, ratings, etc.). Sometimes, we may only have user, product description but not the interactions (e.g., age, product_type). We may have both (e.g., Xbox Live!). And, it is desirable to use the available information in estimating the relevance.

The common approach to solve the problem is based on Matrix Factorisation (MF). MF assumes a common feature (factor) space between the user, item and then these factors are estimated based on the interaction between users-items. In MF, there is no straightforward approach to incorporate native features of user, product (e.g., age, location) unless modelled as users/items.

I will present a unified model that (a) does not compute explicit similarity (like in K-nearest neighbourhood), (b) does not assume common feature space (like in MF) and (c) models users/items with interpretable features (as opposed to hidden), uses all the available information. And, can be used for building large-scale personalised systems in recommendation & search.

Benjamin Habegger, 109Lab: feedback from a start-up experience in collaboration with academia

In this presentation I will talk about my experience as CTO of 109Lab, my first start-up launched with 2 associates, Laurent Assouad and Bastien Marot. The goal of 109Lab was providing a “Big Memory” service helping people sort out, share and print the best of their digital memories (pictures, videos, notes, etc.).  Our target business model was primarily based on selling photo-products based on these memories. Since the beginning of the project, we had the will to work in connection with scientific research, in particular due to my background as a researcher. Therefore, we directly started off with a collaboration with researchers of the LIRIS Laboratory at the INSA of Lyon. The work we lead with the LIRIS was studing how semantically enriching digital memories with contextual information (e.g. exif, spacio-temporal meta-data) could help sorting, organizing and creating stories with these memories. Among this work, we investigated how we could use user-action logs to recommend future actions based on what other user’s had done in the past in a similar context. To help the user sort his memories, clustering and classification techniques  were envisaged. After a two year experience, launching two services Blaboum.com and OhMyToast.com, working on a seemingly hot topic for many people, gaining quite some notoriety within both the research and entreprenship communities in Lyon, we decided to stop for different, likely entangled, reasons. Far from a real failure, this experience has been very rich in lessons on which I am now building upon. This talk will be about this experience, the difficulties in making the mixture take, my view of the possible reasons it did not take for us, feedback on mixing academia with a start-up project but mostly why it was really worth trying and why I am likely to do so again!!

Heimo Gursch, Thoughts on Access Control in Enterprise Recommender Systems

This  talk presents a project that the Know-Center is working on with the Virtual Vehicle and four large German automotive companies.  This project is set in the engineering domain of these automotive companies. The current solution of the information gathering process is unsatisfying for their employees. Due to the diverse and heterogeneous nature of their data services, it is hard for them to find what they need. Strict company policies controlling access to the data make the situation even worse. Enterprise search solutions that are currently available are no satisfying solution, since they are hardly used when introduced in the companies.

The  proposed solution consists of a deep integration of recommender systems and the access control scheme that is applied. This presentation focuses on possible solutions to their problems and why information technology alone cannot solve all the issues. Starting with an overview of the current situation at our partner’s, the presentation will continue with aspects of access control as well as the recommender system.

Nikos Manouselis & Christoph Trattner, Je t’aime… moi non plus: reporting on the opportunities, expectations and challenges of a real academic-industrial collaboration

This presentation will be a live exchange of ideas & arguments, between a representative of a start up working on agricultural information management and discovery, and a representative of academia that has recently completed his PhD and is now leading a young and promising research team.

The two presenters will focus on the case of a recommendation service that is going to be part of a web portal for organic agriculture researchers and educators (called Organic.Edunet), which will help users find relevant educational material and bibliography. They currently develop this as part of an EU-funded initiative but would both be interested to find a way to further sustain this work: the start up by including this to the bundle of services that it offers to the users of its information discovery packages, and the research team by attracting more funding to further explore recommendation technologies.

The start up representative will describe his evergoing, helpless and aimless efforts to include a research activity on recommender systems within the R&D strategy of the company, for the sakes of the good-old-PhD-times. And will explain why this failed.

The academia representative will describe the great things that his research can do to boost the performance of recommendation services in such portals. And why this does-not-work-yet-operationally because he cannot find real usage data that can prove his amazing algorithm outside what can be proven in offline lab experiments using datasets from other domains (like MovieLens and CiteULike).

Both will explain how they started working together in order to design, experimentally test, and deploy the Organic.Edunet recommendation service. And will describe their expectations from this academic-industry collaboration. Then, they will reflect on the challenges they see in such partnerships and how (if) they plan to overcome them.

Thomas Stone, Venture Rounds, NDAs and Toolkits – experiences in applying recommender systems to venture finance

Academic-Industrial Collaborations – I am undertaking research with a venture capital firm called Correlation Ventures and applying information retrieval techniques to areas such as industry classification, peer/competitor identification and matching investors with private companies. Due to the nature of the data I am working with there have been extended delays in terms of getting agreements (NDAs) and receiving datasets (anonymizing, 3rd party consent). Previously I had faced several challenges (availability, privacy, IP issues) in finding suitable industry partners and have a handful experiences to share.

Tools for Researchers – I have had experience working with several different toolkits and libraries for applying recommender systems and machine learning (MyMediaLite, LibFM, RapidMiner, Python (SciPy, NumPy, scikit-learn), SVDFeature). I am also now involved with an open-source project PredictionIO targeted at developers who want to build scalable predictive features in their own applications (content discovery, recommendation, personalization). I would be happy to share my experiences both positive and negative on my experience using these different tools as a research student.

Mark Levy, Item Similarity Revisited

The announcement of the Netflix Prize back in 2006 saw the start of a rush to develop novel methods for preference prediction based on collaborative filtering of ratings, a theme which continues to be pursued to this day in academic circles. The impact of rating prediction methods on industry, however, is unclear. Netflix themselves commented on their tech blog in 2012 that rating predictions formed only a single (and relatively uninfluential) input feature to the ranking models which they actually use to generate recommendations. Meanwhile several other industry players, particularly those whose datasets contain only implicit feedback and not ratings, are known still to use simple item similarity methods as the basis of their recommender systems.

Item similarity methods offer fast computation at recommendation time, and natural explanations of why particular items are being recommended, but they have not been a focus of academic research, except as benchmarks which can apparently be easily beaten by more complex algorithms, perhaps because item similarity tends to give high quality recommendations only when carefully tuned for a particular dataset. An interesting paper from 2012 bucked the trend by introducing Sparse Linear Methods (SLIM), and showing that they easily outperformed more complex preference prediction models for top-N recommendation, but at rather a high computational cost compared to traditional item similarity methods when applied to large datasets.

In this presentation I show experimental results which suggest that a simple relaxation of the problem constraints solved by SLIM can lead to an item similarity method which outperforms model-based algorithms but at reasonable computational cost. I put this in the context of some reflections on the reality of running large-scale industrial recommender systems based on experience at Last.fm and Mendeley.

Why is Mendeley Hosting this Event?

Mendeley employs recommender systems technologies to help researchers organise their research, collaborate with one another and discover new research.  We have an in-house team of R&D engineers and Data Scientists who collaborate with academics in four European funded international research projects.  One of these projects is the TEAM project (http://team-project.tugraz.at/), a research project funded by the European Commission in FP7 in the context of the Industry- Academia Partnerships and Pathways (IAPP) Action within the Marie Curie Programme.  The TEAM project is sponsoring this event, encouraging collaborations to form between academia and industry.

Where is the Event?

The event will take place in Headrooms, 1-2 St John’s Path, London, EC1M 4DD (map).

If you have trouble finding the venue, from Britton Street, look for the Jerusalem Tavern and you should see the arch that leads to St John’s Path (see pictures below).

IMG_20130618_104747IMG_20130618_104756

Advertisements

From → Uncategorized

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: