What does it take to compete in a global arena in which retail and cloud are increasingly intertwined? Domain-specific data science and machine learning for the masses, according to Alibaba.
The war in retail has long ago gone technological. Amazon is the poster child of this transition, paving the way first by taking its business online, then embracing the cloud and offering ever more advanced services for compute and storage to third parties via Amazon Web Services (AWS).
Amazon may be the undisputed leader both in terms of its market share in retail and its cloud offering, but that does not mean the competition just sits around watching. Alibaba, which some see as a Chinese counterpart of Amazon, is inspired by Amazon's success. However, its strategy both in retail and in cloud is diversified, with the two converging on one focal point: data science and machine learning (ML).
Wanli Min, Alibaba's principal data scientist, is a key figure in devising and implementing Alibaba's strategy. ZDNet had a chance to talk with Min about retail in and of the cloud, as well as data science, data pipelines, and ML.
A RETAILER TALE
Alibaba is not really a household name in the US, as the e-retail market there is dominated by Amazon and Walmart with others in pursuit. Recent expansion moves by Amazon and the ensuing demand by Walmart on its associates to move their applications off AWS has peaked the antagonism between them.
Alibaba however is huge in China, and China is huge. This makes Alibaba a force to be reckoned with. Even more so as there is still margin for growth there, both in terms of retail and in terms of cloud. This has not gone unnoticed by global players rushing to China to claim a piece of that pie, but it's clear that Alibaba has the home court advantage there.
Alibaba is not really in the picture for retail in the US. But they are set on changing that, by leveraging new products and data science. Image: Statista
This cuts both ways though, as Alibaba is also aiming to expand beyond its home market. Besides Asia, Alibaba is expanding in the Middle East, the US, and Europe. This brought Min to Paris to investigate partnerships and to advocate, as Alibaba Cloud participated in Viva Technology, the French answer to CeBIT.
Alibaba's record-breaking IPO in 2014 coincided with the launch of Alibaba Cloud. Alibaba looked to Amazon for inspiration there, however its cloud strategy is diversified, reflecting its overall strategy. Alibaba works as an ecosystem of retailers, consisting what it calls an economy.
What this means is that Alibaba wants to be something like a service provider to its retail customers, rather than owning the entire stack like Amazon or Walmart. And now Alibaba wants to leverage its cloud, data, and expertise to become the disciple of digital transformation (DT) for its ecosystem partners.
""The cloud is already accepted, but the question is -- what's next?" says Min. "What can you do with that compute power? Our answer is data intelligence, to provide real-time actionable insights. We are bringing together our cloud, our data and our expertise to facilitate DT via data science."
VERTICAL, VERTICAL, VERTICAL -- VALUE, VALUE, VALUE -- BRAINS, BRAINS, BRAINS
Min refers to Alibaba's recent launch of "Brains": Alibaba domain-specific intelligence solutions for domains such as healthcare, transportation, and manufacturing. This is in stark contrast to AWS, which offers generic infrastructure and tools and lets clients build applications on top of that.
Min explained that the rationale was to diversify from AWS by offering a value-add proposition instead of trying to play catch-up with them. "Convincing clients to go cloud is easy. But we need to convince them to go Alibaba Cloud, and that's where we made a different choice: vertical, vertical, vertical, value, value, value."
This may sound like a reasonable strategy for Alibaba, but it's not an easy one to execute.
First of all, how can you get the expertise for so many domains in one place? For domains like manufacturing and transport, Alibaba leveraged expertise by finding and hiring the right people. But Min says they can't do this for every domain, so the goal is to build strategic partnerships.
"We develop something workable, like a version 1.0, something our partners can start with, and then work with them to build versions 2.0, 3.0 and so on," explains Min. There's just one problem there: how is "something workable" going to compete against specialized solutions that have been developed by a number of domains by now?
"We had our doubts," Min confesses. "Doing this means going against competitors specialized in their area." The advantages of cloud that Alibaba can provide, like elasticity and scaling across geographies, are pretty much a given for these solutions too. Running in the (AWS, Microsoft, Google, etc.) cloud as SaaS means that's not much of a differentiating factor.
So why go for Alibaba? There's always the ecosystem aspect, and Min's answer along these lines, focusing on data science: "We can support clients going into uncharted territory. Our Brains can support you, and you will not be fighting by yourself -- you'll have an army of data scientists on your side."
YOU AND WHAT DATA SCIENCE ARMY?
The numbers there speak volumes. Alibaba has ~37,000 employees, and 20,000 of them are technical. Min is the leader of a cross-functional team of 300 people, including about 50 data scientists, 200 data engineers, and 50 business experts. The data science skill shortage is also felt in China, but Min says they have managed to recruit people from places like Japan, Europe, and the US.
Alibaba's strategy is based on an ecosystem, and it leverages this ecosystem to offer domain specific, data science-based intelligence applications too.
So how do all these people work, and what keeps them busy? Min says when approaching a new domain or problem, they do so in an exploratory fashion, but always with a business-oriented mindset. For example, transportation and logistics was chosen for its potential for impact. Even a single digit improvement for Alibaba partners can result in huge savings.
"There's different stages," says Min. "Initially, nobody knows how much we can do. We investigate feasibility and boundaries -- where it would be possible to break through current barriers. Then we try to accelerate, find better approaches, and invite our partners to co-innovate."
That sounds closely knit, but also labor and time intensive. Does Alibaba consider automating part of this process, or using some sort of framework for this? "Our approach is semi-automated. I don't believe in fully automated data science," says Min. "There is a huge risk there: you may come up with something that does not make sense in the real world.
If you do exploratory work in physics for example, you must make sure that your results are in line with the laws of physics. In business, your results must be in line with business processes. Otherwise you may end up with results that look fine on paper, but not make sense."
There are a number of spurious correlations examples that Min cites there. But isn't the boost in productivity that comes from automating tasks like trying out a multitude of ML models and features tempting? And what does Alibaba do to ensure ML results make sense in the real world?
"We do sanity checks" says Min. "And it is the subject matter experts that do those, not the data scientists. I don't want data scientists involved, I want people with a critical view to do this. They don't know the techniques, but they know the domain, and can tell you whether something makes sense or not.
Yes, it is conceivable that you may get in Go-like situations, where an algorithm may give results that make no sense because you did not think something was possible, but we're not talking about this. We are talking about checking whether your moves are in the board, so to speak. If results comply with the rules, fine, otherwise you have a problem. I see this a lot, this is why I insist."
BLACK BOXES AND DATA PIPELINES
And what about the black box problem with ML? While using ML may give great results, explaining how these result were derived is not always easy. "That's a huge concern," says Min. "Predicting is great, but in the end it's all about actionable insights. Our clients want to know how to improve, which factor to change and why. So we need to have explainable models. I don't like massive data intelligence without paying attention, and our clients often tell us too."
Min's way of dealing with this is by building two models -- a fast one and an explainable one. "We use a black box model to get results fast. Then we try to use a traditional model with explainable structure to approximate our results. As long as we have an explainable model that can approximate results with infinitesimal difference, it's good enough. I'd rather go for an explainable model.
Very often we have a hard time explaining results to customers. If we use the approximate model, it's much easier to sell: this is negative impact, this is positive impact... this matches the expert's experience of the world. They may not be able to quantify it, but they can relate to positive and negative impact."
Min says they build such models that look like sequential step-wise regression to try and mimic and approximate a black box model. But is it always possible to do this when you have features in the thousands? And wow hard is it? For Min, "you need the computational power to run them, but building them is the hardest part.
It takes a while for every new product, as it's a trial and error process. It's even hard to define the problem: we need to account for all input, figure out what kind of output we should expect and so on. We need to decompose the problem in a number of smaller problems, and that requires both technical and business expertise.
For example, my team once came up with what they considered a great solution for a certain problem. But on closer look, that solution depended heavily on a parameter that was vulnerable, as its value came from a sensor that was not 100 percent reliable. So that model was not workable. What happens if that value is missing, or if it's wrong?"
Finally, what kind of architecture and infrastructure does Alibaba use for its data pipeline? Its pipeline is a classic Lambda architecture one, with a streaming layer and a batch layer. It's rather complicated in fact, as Alibaba uses both Flink and Storm for real-time data processing, and in both cases has its own forks that it works with.
Min says the reason has to do with legacy. This is also why the company does not have immediate plans to flatten their architecture to a pure streaming Kappa one, as it has to support existing partners that use Storm.
Min emphasizes that partnerships are the key to Alibaba's strategy for expansion, so in that light that makes sense. Min also claims the "Brain" solutions are tested and reliable and will be competitive against point solutions. It remains to be seen how this strategy pays off for Alibaba, and how much traction it can get.