I attended RecSys'17 for the first time this year in Lake Como. I am not really an ML practitioner and RecSys is an academic conference, so I was largely out of my element. That being said, it was extremely beneficial to be exposed to so much innovation happening in the field of recommendations.

A lot of papers were centered around Deep Learning. Of particular note was Personalizing Session-based Recommendations with Hierarchical Recurrent Neural Networks by Quadrana et al. The paper was really the sequel to Session-based Recommendations with Recurrent Neural Networks. Both papers demonstrate the use of GRU to provide recommendations and are based around a novel way to feed user sessions into an RNN without the use of padding. I also found their ideas around negative sampling quite useful: in a mini-batch, if you randomly sample items you have not seen, you are likely to be picking up a popular item anyway. The fact that the item is not present in the session is indicative in itself. The original paper came with source code, which I did try on a dataset, and without any tuning it provided reasonable recommendations. Of course, it would need a lot of customization before it can be used in a production setting and I’m not sure it can outperform non-DL approaches, but it is a great start! I look forward to trying the ideas described in the new paper.

The industry sessions were also illustrative in seeing how other companies approach recommendation problems. Netflix of course had great presentations, quite relevant to the work we are doing at Tubi, but I also enjoyed the one by EA. In particular, they talked about their approach transitioning between recommenders tuned for cold start (no info on the user) to recommenders tuned when rich data is available after the user has interacted with the service. They also briefly touched on their infrastructure, preaching a motto I strongly believe in: build shared infrastructure that everyone can use.

Lake Como — can’t complain about the setting either