RecSys is one of the most famous and important ACM conferences which captures the cutting-edge of innovation with a special focus on recommender systems. People from both academia and industry join all together to see the advancements in the field. The topics covered are very diverse: algorithms, math, sociology, machine learning, data science, UI/UX, law etc. This year it took place in the beautiful city of Copenhagen in the 3rd week of September.
The event opened with an outstanding keynote by Mireille Hildebrandt, addressing to topic of behaviourism in advertising and micro-targeting, advertising importance, and ending it with an explanation of GDPR in this context. Questions that kept me wondering during this presentation were: “Is the objective of recommender systems to help users find what they are looking for?”, and “Are these recommendations relevant?”. Nowadays, companies keep investing millions in digital ad/marketing. We saw that Procter&Gamble cut $200 million in digital ad spend and it increased its reach by 10%. Likewise, New York Times did the same cut off in Europe and kept growing in ad revenues. So, is it always worth to invest in digital marketing? During the dinner reception, I had a discussion with Prof. Mireille about GDPR, concerns in data privacy and its effect in law and technology. The second keynote by Eszter Hargittai (University of Zurich, Switzerland) referred to inequality in online participation and how differences in online behavior vary by socio-demographic characteristic as well as people’s internet skills.
Deep Learning is taking over the world …
Algorithms were the key concept of most of the papers work presented this year in RecSys. Generally, a traditional algorithm in recommender system is collaborative filtering, where the main assumption of the method is finding user-item similarities in order to recommend items with a similar taste to specific users. However, in this year many papers exposed solutions by using deep neural network architecture like siamese network, two tower DNN, wide and deep NN etc.
An example is Recommending What video to watch next: A multitask ranking system by Google, Inc., describing a large-scale ranking system for next video recommendation. As we already know, Youtube recommendation system follow a two-stage design, 1) candidate generation and 2) ranking system. This paper focuses on ranking, extending wide & deep neural network model architecture by adopting multi-gate Mixture-Of-Experts (MMoE) for multitask learning and also removing selection bias by adding a shallow tower to model architecture. The main idea of this layer is to share experts across all tasks(a task is to predict user behavior related to user utility and a gating layer is already trained for each task) to capture the task differences.
Deep learning techniques in recommender systems have become a first choice of researchers in algorithmic aspect. Many papers building and describing new efficient algorithms, outperforming baselines are being published recently, and as a result, making it difficult to keep track of state-of-the-art at the moment. The best paper awarded “Are we really making such progress? A worrying analysis of recent Recommendation approach” by Maurizio Ferrari Dacrema analysed 18 algorithms published recently in top conferences in terms of reproducibility, efficiency and performance. It turned out that only 7 of them are reproducible. Moreover, 6 of them can be easily outperformed by some heuristic methods based on NN (nearest neighbors) techniques.
Industrial research was another important aspect of this conference, where big companies like LinkedIn, Spotify, Google, IKEA, BBC shared their current research work. Client satisfaction and building a great user experience is at heart of all these companies.
IKEA presented their innovative work with the title “Designer Driven Add-to-Cart Algorithm”. Particularly, they have developed an inspirational shopping experience by recommending products with a notion of style to actual products in view. The recommendation relies on product metadata, user experience and also on inspirational products combined by their designers. They use this information to train a pre-trained ImageNet model. This model suggests to the user similar products in the same style.
Spotify introduced a framework called “multi-bandits” to balance exploration and exploitation, helping them to quickly adapt the home page, based on user preferences. Model training does not take a lot of time thanks to this framework and they train them every single day.
Another pleasing work was presented by BBC. They incorporated in BBC+ app a “host-in-house” recommendation engine, targeting younger audiences for short video clips. However, they have to consider data sensitivity, as a public service broadcaster they must go through GDPR, which recently installed strong implication on the design of system architectures and data privacy.
Bosch is integrating recommender system as SaaS in vehicles, supporting various applications, such a routing (POI or routing), infotainment (music, news or info recommendations), communication, in-vehicle control (seat position , light adjustment and ambient light) and out-vehicle control (alarm systems, heating).
As Olivier Koch (Criteo Labs) mentioned in his blog post for last year conference, also this year the schedule was fully packed and quite a tiny room for Q&A. However, participants could follow up the discussion with presenter during coffee breaks. During coffee breaks there were some demo sessions of build systems. Microsoft published recently a python package for recommender systems using latest personalization algorithms in Azure cloud environment. As Arinti is a Microsoft partner and we work mostly in Microsoft environment, this package will be very useful to help out our costumers in such systems.
In conclusion, RecSys was an outstanding conference and an amazing experience for myself. Moreover, we got to try local Danish dishes, and we also discovered future artists at the karaoke on Tuesday evening :-). I want to thank Arinti for supporting me in this great experience.
This blog was originally posted on https://blog.arinti.be/13th-acm-conference-on-recommender-systems-recsys2019-copenhagen-f2767001be28