Recommender systems have been around for decades, over which time a strong community of academic researchers and industry practitioners has emerged. They invest a copious amount of time and energy into understanding the theory and practice of recommenders.
As the recommender community grew, it got increasingly organised and created a series of conferences named The ACM Conference Series on Recommender Systems. Like many research conferences, attendees share their experiences of what they have discovered by writing their research up in papers and submitting them to the conference for approval. If their work is found to be interesting enough to share at the conference, by a selection of their peers, then the paper is accepted and they are invited to give a presentation about it in person. There’s also a small number of presenters who, due to their great work over the years, are invited to give keynote presentations. This conference series has often enjoyed a healthy mix of both academic researchers and industry practitioners, making it a great place to learn about the challenges from both perspectives.
The series began in Minnesota in 2007 with 16 full papers accepted and presented over 2 days (Figure 1).
Over the past 9 years, the number of full papers submitted has consistently topped 100 and the acceptance rate (an indicator of quality) has begun to stabilise (Figure 2). This year, the conference was held in Vienna with 28 full paper presentations over 5 days, 3 being devoted to the main conference and 2 to a doctoral symposium and workshops.
As the field has grown, the focus of research has changed. It’s interesting to look at the main topics of presentations by looking at the sessions from the conferences to see how trends have come and gone. A session typically organises together around four papers that address a similar topic. We used good old pen and paper to visualise how these topics evolved over the years (Figure 3). Session names are organised in columns and rows, where columns indicate the year in which the session took place and each row groups together similar topics.
The first conference started with three sessions, two being on algorithms and one on user issues. Both of those topics have persisted throughout the conference series although user experience isn’t always pulled out as a separate session. In 2008, the second conference introduced work on recommenders and social networks, another theme that has consistently drawn much attention. From 2011 onwards, two industry sessions were introduced that put the spotlight on work being done in industry. How to evaluate recommenders has been a popular topic, going beyond prediction accuracy and improving metrics. The conference has also devoted much time to exploring emerging themes and novel setups and applications as well as understanding the role that context plays in generating relevant recommendations. More recently the increased work on solving the cold start problem has seen it being pulled out as its own session. The cold start problem occurs when a recommender needs to find a way to start providing relevant recommendations despite having little to no information about the user. A few sessions were just one offs: conversations systems; privacy and security; and innovative preference expression and usage assessments. We have attached a tsv file with the sessions names over the years. If anyone is feeling generous enough to create a neat visualisation of it and share it with us, we’ll make sure to post it!
In total, 300 full papers have been accepted in the series of conferences over the past decade, from Minnesota to Vienna. This is a fantastic resource to explore if you really want to dive deep into the details. If, however, the thought of going through all of these papers isn’t your thing then you may be pleased to know that we’ll be pulling out some of the ideas and techniques and blogging about them in the coming months.