RecSys2020 Highlight Sharing
RecSys 2020 was planned to be held in Brazil, South America, but it had to be changed to an online event due to the COVIOD19. Although holding a conference online wouldn’t allow people to communicate effectively but it also provides convenience for me located in Beijing who is far away from Brazil. Thanks to the efforts of all parties and the application of the software, the overall effect of this participation is higher than my expectations. Here are some highlights from the engineer’s point of view, I hope you learn something from it.
Conference Organization Highlight
Let me briefly introduce the organization of this meeting. I have to say that the organizer worked very hard to make the online conference as good as possible, In order to ensure the normal progress of the meeting, some people keep alive for 24 hours.
- In order to enable participants in all time zones around the world to engage the presentation in a good time slot, speakers are asked to give presentation in two different time periods
- In order for the online meeting process to proceed smoothly, the conference uses Whova software to help the entire process. Whoa includes the meeting schedule, live broadcast platform, collecting questions, forums, making friends, and event organization.
- In order for the participants to gather and discuss freely, the conference uses Gather.town to organize Poster Session and other activities, and everyone can communicate in a virtual space by themselves. In the picture below, I am standing in front of the RECSYS panel in the aisle of this virtual space.
RecSys2020 Highlight
This is my third time attending RecSys. The research progress of recommender systems is not fast, and some strange research directions and repeated research results can often be seen at conferences, but there are always some highlights.
Highlight from Industry Direction
- PURS: Personalized Unexpected Recommender System for Improving User Satisfaction, Pan Li
- Behavior-based Popularity Ranking on Amazon Video, Lakshmi Ramachandran
- Query as Context for Item-to-Item Recommendation, Moumita Bhattacharya
These methods are either verified by the AB test in the industry, or they are simple to implement and can solve some problems.
Highlight from Research Direction
- Counterfactual Learning for Recommender System, Zhenhua Dong
- Debiasing Item-to-Item Recommendations with Small Annotated Datasets, Tobias Schnabel
- A Large-scale Open Dataset for Bandit Algorithms, Yuta Saito
These methods are either novel, broad prospects for the future, or contain open source code or data.
PURS: Personalized Unexpected Recommender System for Improving User Satisfaction
The author Pan Li is a doctoral student at NYU Stern School of Business. This work is the result of cooperation with Alibaba. In order to solve the problem of Filter Bubble in the recommendation system, they proposed PURS. The advantages of this work include
- AB testing on Video Streaming Alibaba-Youku App increased their TVT 4.6%
- It Solved Filter Bubble by Unexpectedness
- It Modeled Unexpectedness by Embedding Similarity
- It implemented Personalization of Exploration
Behavior-based Popularity Ranking on Amazon Video
The author Lakshmi Ramachandran, Applied Scientists of Amazon Video, introduced a work of their Popularity Ranking.
I used to think that Popularity Ranking can use the current statistics of Popularity to complete the ranking. I did not expect that Popularity Ranking can also use machine learning to solve the cold start problem in Popularity Ranking: some new content cannot be ranked in a good position by ordinary Popularity Ranking. , Expose to users.
The author uses data such as some text information of the content itself, previous popularity, and user interaction information to predict the current Popularity, and finally sort by the predicted Popularity. An interesting phenomenon is the length of time the content exists on the platform, which will give new content a high score. The screenshot below mentions the impact of Age Feature (the duration of existence on the platform)
Query as Context for Item-to-Item Recommendation
The author Moumita Bhattacharya is a Senior Applied Scientist at ESTY.COM, an e-commerce website. She introduced ESTY’s technical options for creating related product recommendations. They have 65 million products, which makes their technology choices very different from Amazon Video. The number of videos on Amazon Video may be on the order of tens of thousands.
Related item recommendation is broken into two stages. In the Candidate Set Generation part of the first stage, they use Word2Vector technology to use the content that users search and click as training data, generate Embedding for each content, and then use Faiss: A library for efficient Similarity search, returns a list of items most similar to the current item as Candidate Set.
The second stage uses lightGBM to sort the Candidate Set.
One of the highlights of the work they mentioned was that they used Context to personalize related items. The following figure mentions that the current program can be used as the Context. For example, during Halloween, for a red recommendation related item, is it possible to recommend some items suitable for Halloween and related items instead of only Only the red hat is recommended.
Counterfactual Learning for Recommender System
The author Zhenhua Dong is the Principal Researcher of Huawei’s Noah’s Ark Laboratory. This speech summarized a series of achievements of Huawei in the Counterfactual direction. In order to solve the counterfactual problem, the author proposes Uniform Unbiased Data, which randomly selects content from 1% of traffic and presents it to users, so that every content has the opportunity to collect user feedback. Using Uniform Unbiased Data, the author launched a systematic research and experiment, and achieved a series of results. One of them was the use of Unbiased Data generated by this 1% of traffic, which created a 3% increase in indicators.
Debiasing Item-to-Item Recommendations with Small Annotated Datasets
The author Tobias Schnabel is a researcher from Microsoft Research, who proposes to use small-scale annotation data to improve Item-to-Item recommendations (the same as the related recommendation of an item mentioned earlier). Insufficient intelligence, people come together. Experiments show that this method is indeed effective, which shows that there are a lot of interference in our training data, and it does not fully represent the user’s preferences for all items. The industry may be able to use the previously mentioned Unbiased Data to achieve the effect of the Small Annotated Dataset.
A Large-scale Open Dataset for Bandit Algorithms
The author Yuta Saito is an undergraduate, and in this Workshop showed his work in the direction of RL&Bandits.
One of the core problems of reinforcement learning is how to evaluate the trained reinforcement learning model before going online or conducting experiments, that is, Offline Policy Evaluation. Academia has proposed a lot of methods, many of which are data generated by simulators, but there are few actual data support.
This work provides two sets of clothing shopping behavior data generated through Uniform Rank and Bernoliour Rank, which can be used to evaluate the effects of different Offline Policy Evaluations, and can also be used to develop new policies and use Offline Policy Evaluation to evaluate them. https://github.com/st-tech/zr-obp Open Bandits Pipeline The open source code is of high quality and there are many code comments.
The recommendation system is not like image recognition. Theoretically, recommendation system is a typical application scenario of reinforcement learning. However, it is limited to various engineering constraints and the theory is immature. Reinforcement learning has not been well applied to recommendation systems. Successful application. This area deserves attention.
Summary
This online RecSys experience is unique, and the disadvantage is not as much as I was worried about before. Many participants are also very serious about preparing for speeches. I hope that next year, we can see more highlights from RecSys.
