Celine Xu - The Art Forger and the FBI Agent
Celine Xu is a Senior Data Scientist at Axel Johnson - a Swedish family business founded in the 1870s and now run by 4th and 5th generation family members. They are one of the biggest trading companies in the region, with 22,000 staff and millions of customers each day across a range of different brands.
Celine’s specialism is the application of Artificial Intelligence recommender systems within industry, but is quick to point out that her role is a human and business role and not a scientific or software one. Celine joined us at MTF Örebro in October 2019 and brought her industry and commercial understanding to the innovation labs that brought together AI and robotics with music, dance and creativity.
Celine explains recommendation systems and AI that gets smarter when competing algorithms attempt to outsmart each other. She also talks about the underlying value for industry in participating in something like the MTF Labs.
Music: reCreation by airtone (c) copyright 2019
Licensed under a Creative Commons Attribution (3.0) license.
AI Transcription
SUMMARY KEYWORDS
people, music, data scientist, forger, data, company, business, axel, ai, industry, celine, recommendation engine, neural network, mtf, image recognition, fbi agent, system, problem, create, bias
SPEAKERS
Celine Xu, Andrew Dubber
Andrew Dubber
Hi, I’m Dubber. I’m the director of Music Tech Fest, and this is the MTF podcast. In October last year, we ran the MTF Labs in residence within academia at Örebro University. We were in the robotics lab working with our incredible AI department, as well as with the music school. And we were, as you might expect, crossing the streams wherever possible. Now in the AI labs, we had a range of incredible experts from all over the world, from neural network developers to dancers, sound designers for the automotive industry, to accessibility designers, DJs and beatboxes to robotics engineers. And in amongst that mix, Celine Xu was exactly the same as everyone else in the sense of being not entirely sure what to expect, somewhat surprised to find yourself there and confronted with a group of brilliant people with radically different knowledge, skills and expertise. Now we invited Celine as she’s the lead data scientist for one of Sweden’s biggest retail and manufacturing groups, Axel Johnson, and I wanted to sit down with her to talk about not only who she is and what she does, but how her experience in industry fits into the context of something like Music Tech Fest and the MTF Labs. From an office on the fourth floor of the World Trade Centre in the heart of Stockholm. This is Celine Xu. Enjoy. Celine, thanks so much for being on the podcast. So nice to be here. Thank you. Thank you. You’re a senior data scientist for Axel Johnson, let’s just start with what’s Axel Johnson.
Celine Xu
So, Axel Johnson is a family group they host seven different groups and major business in retail so they have Åhléns, KICKS, Martin & Servera as food holding company. At the same times we have the shareholder of Axel food and Dustin which is online digital stores. At the same time, we also own and company called Novax which is a investment branch and invest in a lot of new business or companies like Flippa co, as a high end fashion a clothing company. And also we also have the training company which is Jean and also we have another group called Axel Johnson international which focus on industry product, which they also have different company within that group, we produce pump and help for example, Arla build their whole yoghurt pipeline,
Andrew Dubber
so everything from all ends, which is like a department store for people to buy clothing and perfume and knives and so on right through to servicing like the restaurant industry. And so you’re a data scientist within that, what do you do,
Celine Xu
I’m responsible for inspiring and supporting the holding company within the group such as Åhléns, Martin & Servera to become more data driven and customer centric and I try to help them deliver and production lines the end to end advanced analytic applications and insights Okay, what does that mean? Okay. So, I have three main roles in my job. First one, of course, the data scientist, which is I translate the business problem into a data model which can be used or code in programming, and then let the machine learn or predict or optimise and produce the result or the insights and then translate that insights back to actionable plan. This is data scientists part the second row is kind of strategic analytical strategist and machine learning specialists. So like the company x food if they have specific use case, for example, tron prediction, they will like to have me as a second opinion to ask whether or not this algorithm is good or not. Or Martin & Servera will have analytical plan in for example, customer domain. And we will help them to formulate and prioritise different user case based on the high value high impact and high visibility. And the third row is I’m also an community leader within the group for data scientist and the sum of data engineer, we try to build this analytic Academy, let people share knowledge because as you know, currently in traditional industry, the like data scientist is alone in the company. So there are at most one or two in the team. So normally we want to have this Forming community to support them. And also, to take advantage of We are a group of family. So we try to share the experience within the whole group to leverage the knowledge.
Andrew Dubber
It sounds like for somebody who works with data and computers, a lot of what you do is about people.
Celine Xu
Yes, of course, fundamentally digitalization of the company is an people game. It’s more about how you change the way people work. It’s not only about algorithm is not only about how you take advantage of the machine learning, at the end, it will be the people to use that machine to help them to generate value or make their life easier.
Andrew Dubber
This, some people say that there isn’t an industry anymore, that isn’t a data industry. It says like all of the things that you do, whether it’s online retail, or whether it’s service to restaurants, or whether it’s finance, or they’re all data now. So is what you do something that could be done pretty much in any industry.
Celine Xu
Yeah, I think so currently, the algorithm people use or the language or programming language people use already exists for a long time. So the basic logic are the same. But it depends on different industry, you or people just to blend in their business logic, business rule, and their understanding about business. That’s make things different.
Andrew Dubber
A lot of people when they talk about data, and particularly AI, the first thing that people think of is Oh, my job is now going to be done by a robot is any of what you do about replacing people?
Celine Xu
No, we actually try to use the machine to free people time to let people do more creative or more important work. So the work will be taken away is kind of boring work. In Axel Johnson group, we actually have the plan, as I said, the analytics Academy, we try to retrain people in the way they need to be adapt. So we don’t want to substitute people we just want to make their life easier and let them to do what they really want to do,
Andrew Dubber
but also the nature of their job could change,
Celine Xu
the nature of the job is changing definitely. And also the job description also changed. There is a company goal in Axel Johnson which is 10 50, which is waiting 10 years, we will change 50 of our current business. So definitely the job will changing. But we will not dumping person, we still hope person can be retrained and a way try to facilitate that.
Andrew Dubber
What’s your story? How did you get to be in Stockholm working for this big company in data science? Where did that start?
Celine Xu
I think very good two stories. First is, I would say data scientist is not actually a scientist. It’s more multi discipline title. So I would not call myself a scientist, I would love to call myself more a businesswoman. Second thing is I would say there was a funny story about Swedish immigrants. Someone said it’s either for work or for love. So my story is for love. So I get married to my husband. So I moved to Sweden. Okay. From London, actually, I was working in a private equity firm in London. Okay.
Andrew Dubber
You’ve been in business consulting, and financial area. And now even though you have the word scientist and your job title, yes. It’s it’s a business position that Yeah,
Celine Xu
yeah, I would say it’s kind of special because I’m came from mathematic background, and I always do analytic. Even when I was in strategic consultant or private equity, I still dealing with data. The different is I didn’t really coding at the moment. But since I moved to Sweden, I realise I can’t speak Swedish. So I need to develop the other kind of language to leverage my knowledge, then the coding tap in and it become the nicest part of my life. Because at the beginning, I will say, I can use the material people produced to prove my hypothesis, but now I can produce that material by me. So, so I would say I’m empowered.
Andrew Dubber
We we like growing up, what do parents do?
Celine Xu
Both my parents are university professors. Okay, and what sort of thing?
Andrew Dubber
That’s it? So,
Celine Xu
no, my mom is computer science professor and my father is a economics professor.
Andrew Dubber
Oh, that’s a was the story told for you, then you It sounds like those things have come perfectly together for you
Celine Xu
sounds like but in the realities, absolutely the opposite. Really. My father wanted me to study economy. And my mother wanted me to study computer science. And when I choose my major in university, I actually say no to both that I choose maths because I want to be mutual. And after that, they have big fight saying, for master degree, what I should study that I chose to work first. So at the end, everything come together. But the past is not that
Andrew Dubber
straightforward. No, but they must be pleased with words in devout.
Celine Xu
Yeah, yeah. They’re quite happy.
Andrew Dubber
So let’s talk about recommendation systems. Yeah. Because that’s something that I know that you’re very involved with. I guess my first question would be, why do so many businesses need recommendation systems?
Celine Xu
Okay, 90% of data in a word created after 2010. And all this abundance, present a big problem, the Paradox of Choice, because we have so many choice, and we need to spend too much time try to pick one. And sometimes we try so hard, but at the end, we actually pick something wrong. And the recommendation engine is actually use machine learning technology to help company go over all the possible options and learn what we are as a customer like and recommend the options we love best. So this machine or system providers an option, having the abundance of options, at the same time have a certainty in our decision.
Andrew Dubber
It sounds like the assumption is that choices are difficult. But I don’t experience that when I get to a shop, for instance, and I want to buy a sandwich. It might take me a moment. But I don’t need something to say you should have a ham sandwich, not a chicken sandwich. But it feels like that’s where all of the recommendation engines are trying to do. Is it more complicated than that?
Celine Xu
Yeah, definitely. Yes, I would say, Yeah, I agree people don’t really have the problem to decide what they want. But they do have problem to a well the other possibilities. So they’re different metric to measure whether or not is a good recommendation engine. One, yes, is accurate, which is recommend something you definitely want. But at the same time, you also need to think something they never think about to inspire them. For example, if you never know Vietnam food, whether or not you will order it, and maybe after you taste some of that you will fall in love in some kind of dishes. One of the things is from that perspective, from recommendation engine,
Andrew Dubber
because recommendation engines that I can think of for instance, like, if I buy something on Amazon instance, people who like this, they also like that, and it doesn’t seem to stray very far from the things that I’ve already said that I like music the same. When I was growing up, I would go to record stores and I would buy a record and the person behind the counter would say, you don’t want that record. That’s a terrible record. You want this record, which you’ve never heard before. You don’t like this music yet. But this is objectively good music, take it away, listen to it, listen to it again, until you like it. I still want that kind of recommendation can can an algorithm do that? And say no, no, your choices are bad. Here are some good ones.
Celine Xu
I will say the algorithm itself will not tell you which is good, which is bad. It’s just a based on different business rule based on the company’s logic and give the ranking of the choice you could have and promote the one more close to your choice. Or more close to the company’s target.
Andrew Dubber
Yeah, the things that the company is trying to sell rather than the things that might
Celine Xu
sometime Yes, yes, sometimes Yes. Unfortunately. That’s why I will say fundamentally, the recommendation engine is just m predictive automatic scoring system, okay. And it’s only based on historic method. So they have their own disadvantages for example, because they only based on historical record, if you have drastic change of your taste. They probably will not predict. Right,
Andrew Dubber
right. And I guess if I’m not typical of the customers as well, like, if you’ve got a shop where, I don’t know, let’s say 90% of the customers are really into chicken sandwiches. Yeah. And I just don’t like to consider which is likely to recommend me from the chicken line. So is there this sort of bias that is built into systems like that? Or can it be? Is there such a thing as neutrality? I guess is my question.
Celine Xu
Unfortunately, from statistical point of view, know, everything have bias is just sometimes people do not realise their biases, to example for for you is yes, in some of our business, we have 90% of the loyalty customer are female. So when we recommend something to male customer, we do have some problem, for example, force, eyelash to a male, and we did hear the feedback is is kind of intimidated. So if you don’t add extra layer to regulate the recommendations, you got this bias. But we have some other way to make it better. For example, we can just take away some sensitive product, or we can set up some category only recommend to men, but at the same time we add the other kind of bias in the recommendation system, there is a chance there. Some men do want the false eyelash. So I will say sometimes when you see the bias is built in, it seems the problem of algorithm, but it’s not. It’s actually problem of input data. So if you have 90% of the shopping like histories from women, it’s not because the run of the algorithm is just because of the input data,
Andrew Dubber
right? But at the same time, you don’t want to go through all of your products and go This one is only for women, this one is only for men, you know, because of the problems that that? Yeah,
Celine Xu
yeah, everything have like the advantage and disadvantage. Normally, I would say, in different companies, they have tag, or at least have the department tag for certain product. So we do can find the tag, which one is for women, which one is for men. But we also debate within the company says like sometimes men buy the gift of women who do have the history for the women product. If you only recommend men’s product, maybe it’s not good for them. They also want some inspiration. So maybe the solution could be just a takeaway of really sensitive, productive, but still trust or believe the data?
Andrew Dubber
is more data the answer to all of these sorts of questions, yes or no?
Celine Xu
From algorithm perspective, the more data you collect, the more accurate you have. But at the same time, the more data you have, the more bias you have. Okay, so it depends
Andrew Dubber
on what kind of problem you want to solve. Is this harder in a place like Europe, where there are all these laws about data privacy and protection and those sorts of things? There is, for instance, in China, where those laws don’t exist, where there is such a massively increased amount of personal data so that recommendations can be much more targeted, much more accurate, or do you not experience that?
Celine Xu
It definitely have some impact. For example, in China, you can get more data source, not only from purchasing history, you you have the online data, if you don’t have regulation to minimise the online shopping IP address, then you can easily link the online IP to your store purchase history, which make the only channel really food and here in Europe, you have more regulation. So sometimes it’s really hard to matching or people don’t even want you to match. So when you lose that part of information, yes, their condition will not as accurate as that one.
Andrew Dubber
So what are the uses of something like artificial intelligence for one of a better term would you use in an industry like this not just recommendation engines? Are there other uses of artificial intelligence,
Celine Xu
of course, it depends on how you core artificial intelligence, for example. Now one of the things I also focus on is time series prediction. So basically, is all kinds of prediction, like price prediction, demand forecasting. And those things are really used in the industry and also people planning for example, the workforce planning for department store for Black Friday, how you increase the workforce within the store and how you reduce that it’s also prediction, and I would say two big part currently increase is one image recognition and natural language processing, image recognition you You can tell the story from China like now, Ali pay Alibaba, you can just take your picture and enable you to payment, instead of you the fingerprint, you use your facial at the same time. So a lot of law enforcement, as I know, Alibaba have really successful example, to help the traffic regulation to detect the speed tickets, so people can get away of that, even though they hide the the number of the car, they can use the image recognition to find them. augmented reality, for example, if you use your iPhone, now, you actually can see where you are going, there is a big arrow in your phone. And it’s helping you to find your way. And natural language processing. Also really interesting. For example, currently, you use your phone, and you just say like a serial, remind me blah, blah, blah, and it’s directly put your phone call appointment in the calendar. Also, one of our company have this experiment is Martin & Servera. They use Google Home to let the chef in the restaurant real time order, what material they need to get, right. And it reduced them to typing and it’s really smooth. So it’s quite
Andrew Dubber
nice. Okay, so I guess my my big question is, how did you can make this with coming to Music Tech Fest? And Örebro? What was that about?
Celine Xu
Okay, I will say at the beginning, I was intrigued because the name name is Music Tech Fest role. So, it’s interesting. It’s exciting, because it’s fastest party, and then I get there. And I was so surprised because it’s totally different of my expectation. It’s really challenging, really innovating, and a really intense. Afterward, I understand the mission for Music Tech Fest is actually create this three days platform to let people innovate the new format of creating the music, which is fantastic, I will say, we had a lot of really nice experience, how to use the dance to create the music, how to use the reaction of the music to get the lyric of the drum, and also how to use the deep learning which is again system to natural generating the music, which actually helps human to generate the music in a large scale, which single human cannot
Andrew Dubber
write. So what did you do there? What was what was your role within that context?
Celine Xu
Yeah. So because I know really less about how to compose the music, so I use a label help form the AI Of course. So I try to use GAN system to transfer the music style to a specific song. For example, I find two songs, one is Linkin Park, which is light rock. And the other one is more classic singing. And I try to apply the rock style to that thing by using this algorithm.
Andrew Dubber
When you say GAN system. Can you explain what you mean by that?
Celine Xu
Yeah, the name is generative adversarial networks. So in general, it’s an neural network and the use for unsupervised machine learning. And it made up for two competing models competition with each other. The first one could generator network The second one was discriminator network. So if we make an analogy of that to explain is if we have a masterpiece from more net, say waterlily and there is an forger which one to create duplicate one to sell. And this forger needs to learn how original painter monete painted this water lily. At the same time there is the FBI agent investigator tried to capture this Forge. So he wanted to have the second guess of the route this forger use which enable this FBI agent to detect which one is future. So if we map this to this GAN system, the generator network will be the forger and this network will learn So called the distribution of the class. At the same time, the investigator from FBI will be the discriminator. And it will learn the boundary of those class and then together will make the system more advanced to produce the fake. Okay. panting
Andrew Dubber
because the the forger has to fall. The FBI agent, yes. And so as to get continually better. Yes. Understand. So this is when you say adversarial network, it’s it’s basically two systems that are competing against each other to make both of them better. Yes. Okay. So you brought that to MTF to put different genres of music against each other to create something new? Yeah. What did you bring away from that back to industry?
Celine Xu
Okay, at first, I was so exciting just into the music creation. So I would say first thing I learned is I have better knowledge about game system, or in general neural network, how it capable was. So it got me a lot of new idea how I can apply that things in different area. At the same time, I will say meets a lot of excellent people to open my mind about all the other possibilities, like how AI actually helped creating not only music, but the other arts. The first things I will say, I actually get a lot of impression about how AI help disadvantaged people to pursue their dream, because we have some friends have some huge vintage disadvantage to use the tool, or they’re blind or the hand of them is not that good. So because of that, they actually can develop a system to help them to make their digital life really easier. which make me first a few really proud or empathy. Yeah,
Andrew Dubber
at the same time, it’s just say like, the other side of AI, it’s not like get rid of other people is actually helping people. That’s that’s really interesting, because one of the things that I think is most people notice about MTF is that there are so many different people from different backgrounds, different disciplines, different, like artists, and scientists and academics and business people and those sorts of things. And one of the things that we like about it, is that it brings those people together, which is great in the context, but I’m always curious about how much of that goes back to people’s businesses. And how much of that thinking and though that sort of interacting with different kinds of thinking comes into now when you when you’re not at MTF, does that make any difference?
Celine Xu
Yes, this is only like a neural network. So I don’t know why but it’s a black box. It definitely affected my my my daily job of daily life first, as I said, is open. Other kinds of possibility to let me know more new things. But if you say like go back to the industry, I would say more from inspiration perspective for my self for my life changes a lot. I open more to different kinds of people and talk about different things. I started to learn a new music skills I’m trying to learn flute, Chinese flute. So it’s straightforward instead of sideways. At the same time, have more passion about deep learning and image recognition, actually, so I would say probably it’s take time, but I will try to bring those value back in my daily work. I think it sounds like you were
Andrew Dubber
the FBI agent.
Celine Xu
Oh, no. Hopefully, no, I want to be a forger. Okay.
Andrew Dubber
Celine, thank you so much for your time today. Thank you. That’s Celine Xu. And I guess that makes me the FBI agent. So that’s case closed for this week. Hope you enjoyed. And remember, we’re going to be running the five day Industry Commons ecosystem ice labs in Mannheim in April, all about sound design for urban and industrial applications in the age of AI, IoT systems and blockchain. If you want more information about that, jump on the newsletter at Music Tech fest.net slash newsletter. As always, please know that you can subscribe to this podcast anywhere that you might listen to any other podcast and of course you can share like rate review and go through the growing back catalogue of MTF podcast interviews. I’m Andrew Dubber. Hope you have a great week and we’ll talk soon Cheers.
