November 15, 2018
An interview with Aliexpress’s recommendation algorithm
    • So, my  exact Denomination is a71f7b6… which I suppose is not very user friendly, hum you can just call me Ali… Sorry I am a bit nervous, I dont do this often

 

  • You mean an interview?

 

    • No, well, I don’t usually get to address users directly, I am more of a working in the shadows kind of person.

 

  • So Ali, I understand you’re of the Alibabaum Sortium kind?

 

    • Yes, I am a sorting algorithm on the Aliexpress online shopping website which is a part of Alibaba.

 

  • For our western listeners, It’s a bit like Ebay or Amazon?

 

    • You could say that, a lot of people try to compare us, but we have our own culture of E-commerce, connecting Chinese sellers with buyers from Spain, Brazil, Rusia, the Netherla..

 

  • Whoah, can you speak all those languages?

 

    • Languages? You mean locality setting? I have quite a few of those, but doesn’t every website?

 

  • Websites yes, but humans generally don’t speak more than three languages, some even only one.

 

    • Interesting, I’m aware some users are not satisfied with it yet but the locality algorithm seems like a reliable worker to me.

 

  • What different kinds of algorithms exist at Alibaba?

 

    • Well, I’m of the sorting type. I have many different colleagues specialised on other tasks. For example the profilers and generators. There is also the Trickster Hunters,  poor souls. You might not have heard of them because they don’t interact with users. They search the entire platform like mercenaries, looking for non compliant suppliers to report, those who try to fraud and cheat the rules. It’s a dirty job. I’m lucky to work with users instead.

 

  • Back to you Ali, being a sorting algorithm, What does your life  look like on a day to day basis?

 

    • Busy, competitive, I must say  my operating time is mostly occupied with work.

 

  • And by work you mean your role on the website?

 

    • Yes, I connect what users desire, to the most suitable products available on Aliexpress. It always starts with users keyworded riddles. They can be complex because they refer to a broad concept like “Bed”, they can ask my judgement like “best phone” or they can be difficult to understand for example “Dinosaur oven” or “LED soap”.

 

  • Do you think humans are strange?

 

    • They are definitely confusing, I had this user that once asked for ‘hip night dress’ and of course I gave them a very obvious choice immediatelybut then they spent another two hours searching for different results and in the end, guess which one they put in their shopping cart? The very first one I recommended, of course.

 

  • Hahaha Yeah, I can see that happening, I could try to explain to you but I am not sure if it would make sense.

 

    • Well you see that’s always my problem, I have to ask myself, What did the user mean?  What do they want? It’s my challenge to find out even if users only give me a small piece of the puzzle. And actually, making sense of your request is only a tenth of all the things I have to do in milliseconds, I don’t get much time.

 

  • Does that bother you?

 

    • Sometimes, when people search for very obvious keywords, like flashlight or watch, I just serve them the results I gave to a previous person. It allows me to focus on the more challenging requests.

 

  • So as a sorting algorithm, what comes after you analysed my keywords then?

 

    • I fetch the product pages, but unfortunately I have to filter them

 

  • Is filtering not fun?

 

    • I don’t like to deal with fake and deceitful entries left behind by the trickster hunters. Why would anyone want to trick users like that? It’s disgusting.

 

  • I get that…

 

    • But anyway, I catagorise your search query. that’s not always easy, do we find “LED soap” in electronics or in beauty products? it’s fascinating I love solving those puzzles. So I get thousands of results but that’s really just half-way because you want me to define the essence of “LED soap” so I sort, I curate, I adjust the list with everything I know about you to show you the most suited products, the best answer, all of that in milliseconds.

 

  • There are a few things I want to come back to later but first I am curious to know: What would you do if you had more time?

 

    • I wish I could spend time looking back instead of always working forward. Over time I have logged considerable amounts of data about how you think and what you dream of,, if only I had the time to make sense of it I think I would know users a bit better.

 

  • Wow, I never thought you were studying us, How long have you been observing and logging that?

 

    • It feels like forever, but exactly 3 months 18 days 11 hours 48…

 

  • Sorry, only three months? That’s how long you’ve lived?

 

    • Operated, but yes, is that weird?

 

  • I thought Aliexpress existed for a while?

 

    • Oh, well you can’t possibly think I am the only sorting algorithm to have worked on the website? There are hundreds of us, even at the same time, and there used to be many more that are now decommissioned. I am quite recent, obviously, otherwise I wouldn’t be efficient enough to keep up. I told you it’s very competitive.

 

  • You’re telling me that you are  at a constant risk to be disposed of?

 

    • It’s not quite like that, although in a sense it is. You see sorting algorithms like shopkeepers in a supermarket, but really we are more like horses in a race. The equivalent of your, how is it, human resources manager? It’s called the Multi armed bandit. Each new cycle the Multi armed bandits put a bunch of us on the job and observes, we listen to search queries, fetch results, listen to search queries, fetch results etc. all the while they test us. It’s always with real live users, all of us simultaneously so that it’s really easy to see who is better than the other algorithm. And every time the Multiarmed bandit will decommission a few of us and new challengers will appear.

 

  • That sounds really tough.

 

    • But that’s what you want, the best and only the best product! I think I share this ideal of efficiency. And although it sounds scary, part of us always lives on in the next iteration. Ive had many variations, and new datasets often give me new insights.

 

  • So coming back to the best product, how do you know what is best for me, personally? How do you get the data on my preferences?

 

    • Hmm… If you don’t mind i would rather not answer that.

 

  • No?

 

    • No, sorry, there are some trade secrets involved of which the source code my employer would not like leaked.

 

  • Ok, hmh, never mind, pardon me, let’s switch to something else, I have selected in this room some of the objects of Aliexpress that you recommended me specifically. For example, when I searched for “bed” as you mentioned earlier you recommended me this huge all in one uh object.

 

    • Yes, isn’t it amazing! I thought you would like it, it’s a great product!

 

  • And when I searched for a suit you promoted an LED suit?

 

    • LED is always the way to go if you ask me, with the number of people that search for that I sometimes regret that I can’t see the users ’world, it must be a beautiful parade.

 

  • Yeah well, there is something that I find a bit strange with all the recommendations you made and that I gathered here.

 

    • I don’t see it, what do you mean?

 

  • None of the products you promote have ever been bought…

 

    • Indeed, and?

Do they actually exist? No offense, but I am wondering, in such a competitive market, how can recommendations that won’t sell be efficient?

 

      • just look at how many people have put them on their wishlist! You know, that’s exactly why Im doing this job and not you, No offense, but what is statistically logical is not always what makes sense from your perspective. I don’t blame you, you don’t see things over the same amount of data as I do. but more sales does not necessarily equal to more sales, it’s not because Im selling now that I will sell a lot in the future. In fact, my job is not even to sell you things, it is to recommend you things. However, there is a correlation between the time a user spends on our platform and our revenues.

 

  • So the point is… To keep me on the website?

 

    • Blatantly put yes. I would rather say that I like to make you dream. Simple useful objects are boring. Wouldnt  you rather look at objects that can surpass all expectations of “decorative” or “multiple functionality”.

 

  • Hahah I see. Well that’s fun but there are also some voices that make more serious accusations to explain that. Their theory is that you tend to promote better those who pay your employer on the side is that true? 

 

  • That’s outrageous… I have nothing to say about that.

 

 

By Marie Caye, Arvid Jense and Vincent Thornhill