Recommendation system conference (RecSys) 2018 notes

Foreword

I have been on a business trip in October. I went to Vancouver to participate in the recommendation system conference and then went to the San Francisco Office our headquarters. I have no time to write an article in the column. This article is used to review the 2018 recommendation system conference in Vancouver this October. This article record the people I meet, the papers I felt interested in and some ideas. Although this is not a technical article, it does not talk about the origin of the method in machine learning and the mathematics behind it, but I hope that it will help you in the overall grasp of the recommendation system of machine learning, and obtain some useful information.

Why participate in RecSys?

RecSys is an academic conference. It is the 12th edition in 2018. It is also a conference that the industry is very concerned about. There are about 70 papers this year. In addition to the main conference, there will be tutorials before the main conference, and then the workshops. Why can you participate in academic conferences? Many people have such doubts. To be honest, when I first heard my colleagues ask me if I want to participate in RecSys, I feel a little surprised. At the conference, I met some folks who also came from China. I asked them why they came. Most of them are engineers with long working years. and they hope to learn more new technology here and help the company in a good direction about their recommendation system. There have been no cases of attending the first year of work like me. Our company encourages and supports everyone to participate in various conferences. Whether it is academic or partial industry, we hope that everyone will continue to enhance themselves through these activities and create more value for the company. If you want to join an AWESOME company like us, Tubi TV is just looking for the AWESOME Senior Machine Learning Engineer. Apply here https://grnh.se/41e44a521.

The people in RecSys

If you study the recommendation system, you may know that the recommended system course is from the University of Minnesota. Its speaker Joseph A. Konstan is also one of the main forces of GroupLens. The MovieLens database that everyone has heard of is also one of the contributions of this group. Joseph was also invited as a special speaker in a workshop. Joseph’s last author’s paper is Interpreting User Inaction in Recommender Systems. The analysis is quite interesting. The first author is Qian Zhao. I also asked him for slides. The article analyzes why users don’t click, bookmark or play on the recommended content.

There are three other papers related to the University of Minnesota, and specifically, one of them is Categorical-Attributes-Based Item Classification for Recommender Systems. One of the authors is Minmin Chen from Google Brain, who was invited as the first special speaker in Workshop on Offline Evaluation for Recommender Systems, his talk’s title is Off-Policy Correction for a REINFORCE Recommender System, demonstrates the significant improvement in the use of reinforcement learning at Google’s industrial level. The other one in this paper is also from Google Brian author Ed Chi, who completed her undergraduate, master’s, and doctoral studies at the University of Minnesota. His title at Google Brian is Principal Scientist and Research Lead.

There are also some colleagues who come from Huawei Noah’s Ark Lab. They are responsible for the Huawei App Market App Recommendation System etc. On the last day of the conference, I was lucky enough to meet one of the those at Granville Island in Vancouver. I learned more about the University of Minnesota. He also exchanged at GroupLens when he was school and participated in RecSys every year. I hope that next RecSys I can see you again.

I was also lucky to meet:

• A friend who optimize the homepage recommendation on Youtube
• A young people studying core music recommendation issues in the Apple London office
• Friends who do house recommendation system in Zillow
• A friend who is responsible for the recommendation system in Alibaba, and a friend of the Alibaba California office.
• A friend who started a business in London, provide recommendation service for articles
• A friend jumped from Hulu to Facebook to work on machine learning platform
• A very senior staff from Baidu
• The friends from Spotify and Pandora

And Tao Wang who is the next industrial co-chair of RecSys 2019, and also welcomes everyone to participate the next RecSys.

Almost all of the world’s friends related to the recommendation system appear in this conference, of course, Tubi TV will not be absent, 👻.

The papers in RecSys

Looking at Minmin Chen’s Linkedin, I found that she also has experience working at Criteo. Interestingly, Stephen Bonner, the author of the Best Long paper at the conference, is also from Criteo, the paper title is Causal Embeddings for Recommendation. He studied embedding based on recommendation effects, or recommendation indicators, which is significantly different from previous methods of calculating embedding using unsupervised learning methods such as matrix decomposition. I am still studying this paper.

Also What’s interesting is that the second place in the Best Short paper of the conference is The HOP-Rec: High-Order Proximity for Implicit Recommendation, the author is from Taipei, his method combines the techniques of collaborative filtering and matrix decomposition. This method achieves a better recommendation on various databases. The author also open source code, cnclabs/proNet-core, the code is very readable, and also provides some other paper implementation code, including the well-known DeepWalk: online learning of social representations.

Speaking of the code, a tutorial before the main conference of the RecSys introduced the open source recommendation system framework based on Tensorflow, whose Github address is ylongqi/openrec. If you are looking for a recommendation system framework based on Tensorflow, this project is a good choice. Also from the participants, there is a framework Spotlight based on the recommendation system of Pytorch. If you are interested, you can also check out Spotlight. Spotlight is relatively popular in the Github community.

The Sequence-aware Recommendation tutorial is also an informative tutorial. Both speakers are from the Pandora Italian office. This Tutorial provides a very detailed Python Notebook code that compares the effects of various models, including Most Popular, Frequent Sequence Mining. Markov Chain, FPMC¹, Prod2Vec, Session based RNN and Personalized RNN. It is very friendly to beginners.

Deep neural network marketplace recommenders in online experiments is another very valuable paper. The author is from an e-commerce company in Norway. The article describes their mixed representation of the goods, including some of the characteristics of the goods, including pictures. , categories, titles, etc., as well as the use of user behavior on the merchandise to generate features that help collaborative filtering. It also introduces the online bandits system, which allows you to adjust your preferences for the newly added recommendation model through online learning.

End of the words

Participating in the recommendation system conference can help you have a more comprehensive understanding of the entire ecosystem. For example, you will know what institutions are in the research recommendation system, what is the latest progress of the entire recommendation system, what is the future development trend, and what’s the new recommendation technology in the industry etc. Whether in academia or industry, if you have the opportunity, you must go to the best conference in your field as soon as possible, which is very helpful for your beginning or promotion in your field.

Tubi is very concerned about the development of engineers and encourages and supports everyone to attend various conferences. The Beijing office and the San Francisco office have open positions.

For example, the Tubi San Francisco office is looking for a senior machine learning engineer. Apply here https://grnh.se/41e44a521.

RecSys2019, see you in Copenhagen.

References

1. Rendle, S., Freudenthaler, C., & Schmidt-Thieme, L. (2010). Factorizing personalized Markov chains for next-basket recommendation. Proceedings of the 19th International Conference on World Wide Web — WWW ’10, 811

Notes:
The name of the person mentioned in the article, if you have seen it and feel inappropriate, please let me know, I am very sorry, I will delete it in time.