Give Users the Wheel

**Kyle Polich** (0:06)
Welcome to Data Skeptic, a podcast exploring the methods, use cases and consequences of recommender systems.
Welcome to Data Skeptic Recommender Systems.
Are you one of those people that was upset when Netflix went from five stars to just thumbs up, thumbs down? I'm told that's an episode of the Big Bang Theory. They didn't ask me clearly, cause I don't approve of it. Doesn't the larger cardinality give me more information gain? Can I not express myself more with five than with two? I guess I'm not trusted to do that. But it makes me thumbs up and thumbs down in weird ways, or maybe hesitate. Does thumbs down mean I don't like it, or I don't want it right now, or I don't like this particular one, but show me more like this in the future, versus I dislike the whole category? Today we're going to discuss the paper Give Users the Wheel. This is a proposed hybrid approach, and in my opinion, a bit of a vision for next generation recommender systems that I hope take off. Let's get into it.

**Fuyuan Lyu** (1:09)
My name is Fuyuan Lyu. I'm currently a fourth final year PhD in McGill and Mila Quebec AI Institute.

**Kyle Polich** (1:16)
Can you share a few details on what you're studying in your PhD?

**Fuyuan Lyu** (1:20)
Mostly I was studying on recommendation system and how things are utilized in recommendation systems. But now as I'm moving more and more senior, I've seen more and more recommendation at one specific area, a domain-specific area and how my general technique can be applied beyond recommendations.

**Kyle Polich** (1:37)
How did you first take an interest in recommendation systems?

**Fuyuan Lyu** (1:41)
It's very coincident because I joined an internship the second year of my PhD.
So actually everyone was using the recommendation system. So it's sort of the first thing that you interact with the technical sectors. But the formal version of why I get started in recommendation system was the second year of my PhD, I joined an internship and that completely shaped my interest.

**Kyle Polich** (2:04)
Could you share a few details on it? What was the project you were brought on?

**Fuyuan Lyu** (2:08)
The project I was brought on is more doing CTR predictions. Because previously, I sort of shift my research interest. At the beginning of my PhD, I was doing new architecture search and LML. And at the time of 2019 and 2020, when people just realized that transformers are going to unify everything and people just shift to transformers. And I feel a bit down at the beginning of the internship because I was sort of impacted by that direction. But then I realized that in a lot of real world applications, such as recommendation, the real power of LML on neural architecture search does not lie on how you search on the model space, but how you search on the feature space.
Because the feature itself are very domain specific and there is no unified way of doing that. And how do we provide a systematic way to alleviate human beings from the repetitive effort of feature engineering, because previously people realized on feature engineering a lot, experts, experts' choice, this kind of stuff to make the performance do good and make their model performs good. But then I realized that as the system are evolving and as the scenario are getting more and more complex in our real life, we sort of need a systematic way to automate some part of things. And my first work was trying to automate one of the, I think the key in feature interaction or sorry, the key in recommendation or CTR prediction, which is the key name in advertisement sort of is feature interaction.

**Kyle Polich** (3:32)
So the primary paper I invited you on to discuss is titled Give Users the Wheel. What wheel are you giving them?

**Fuyuan Lyu** (3:40)
So it's all come from my personal experience because I was a heavy Bilibili user. Bilibili was basically Chinese version of YouTube. So in Bilibili, basically I have two mode. So it's really personal. So the one mode is when I wake up in the morning, or like say in the afternoon, I was trying to be more informed of what's happening in the world. Like what's the latest news in the technology sectors, or what happened with the global event? What are the major news? But after 8 p.m., I will sort of shift into another mode, which I was just interested in dogs and cats and funny videos. So because my interest was quite split across time zone, so I was thinking, so I wouldn't explicitly say I disliked this video. Like there is a small button also on B2B2, as I think there is basically any recommendation platform saying that I don't want to see these models.