Artificial Intelligence and Business Strategy
In collaboration withBCG
Huiming Qu didn’t plan to work in data science, a nascent field at the time she was pursuing a Ph.D. in computer science, but one course in data mining changed all of that. She started her career in the research department at IBM, transitioned to a 50-person startup, spent some time in the financial services industry, and today leads data science and machine learning in the marketing and online functions at The Home Depot.
In Season 2, Episode 6, of the Me, Myself, and AI podcast, Huiming explains the similarities and differences between her previous experiences and her current role, in which she is tasked with helping customers more easily find the products and services they need as they embark on home improvement projects. (And who hasn’t started at least one of those since the COVID-19 pandemic shifted many of us to working from home?) She also outlines some of the challenges of managing a data set of over 2 million product SKUs and getting pilot programs to market quickly, and she explains why she champions the need for cross-functional teams to execute complex technology projects.
Read more about our show and follow along with the series at https://sloanreview.mit.edu/aipodcast.
Sam Ransbotham: When you think about tackling complex home improvement projects, The Home Depot likely springs to mind. But what complex AI and ML problems does The Home Depot face while helping you with your projects? Find out today when we talk with Huiming Qu, senior director of data science at The Home Depot.
Welcome to Me, Myself, and AI, a podcast on artificial intelligence in business. Each episode, we introduce you to someone innovating with AI. I’m Sam Ransbotham, professor of information systems at Boston College. I’m also the guest editor for the AI and Business Strategy Big Ideas program at MIT Sloan Management Review.
Shervin Khodabandeh: And I’m Shervin Khodabandeh, senior partner with BCG, and I colead BCG’s AI practice in North America. Together, MIT SMR and BCG have been researching AI for five years, interviewing hundreds of practitioners and surveying thousands of companies on what it takes to build and to deploy and scale AI capabilities and really transform the way organizations operate.
Sam Ransbotham: Today we’re talking with Huiming Qu. She’s the senior director of data science and machine learning, products, marketing, and online, at The Home Depot. Huiming, thanks for joining us — from my hometown, actually. Welcome.
Huiming Qu: Glad to be here. Thanks, Sam.
Sam Ransbotham: Can you tell us a bit about your role at Home Depot?
Huiming Qu: So I support this awesome team that has data science and products for overall online and marketing. There [are] many challenging problems that we are solving for — improving the digital experience for our customers — so I’m super excited. I’m not a DIY person, so every day is a learning experience for me as well.
Sam Ransbotham: Customers need help with so many different types of projects. It’s overwhelming to think about how you would narrow down to meet their needs for a specific project. How is The Home Depot using AI to help with those projects?
Huiming Qu: That’s a question our teams continuously reflect on. These are very specialized categories, and sometimes we need to have other business partners involved as well. There is a particular domain knowledge about appliances or flooring, or even plumbing [or] electricians. When we have machine learning algorithms, if we have enough data, we can typically solve a lot of these problems, but a lot of times we don’t have enough data for the niche problems that we’re solving. When we’re going down to the detail of that specific department, in that specific category in plumbing — [such as] PVC pipe — how do we do recommendations the best way? I think a lot of our merchant expertise knows what should go into the recommendation for that particular product, but we have over 2 million products online. How we train the machine correctly, to serve that in a scalable way, is really the key. First of all, we need to make sure we are solving the actual customer pain point and really aligning around data scientists, user experience, product engineering — this really cross-functional team, aligning the goal together.
One example is that we have this project guide recommendation, and we also serve a snippet of the difficulty level of that project. We also serve algorithms in real time identifying, potentially, what is the project you’re working on. If you are searching for “mirror,” [and] you are searching for some of the tools and hooks, then we think you probably need some guide about installing a mirror.
These are things that our customers are facing every day. Especially when we’re at home — now a lot more than before — literally, every day you can think about things that could be improved. We definitely feel the responsibility to help our customers to get the help they need; even just when they search, we need to provide that specific product they’re looking for.
Sam Ransbotham: Shervin and I were just literally talking about how we’re sitting around at home and seeing more things that need to be done now that we’re home more often. Lots of what you’ve described, I might call sort of episodic — like, someone’s at a search and they’re trying to find something and they’re doing something specific. But you have a larger relationship with customers. The search process might be improving an existing way of searching, but can you tell us a little bit about what you’re trying to do with multiple searches and longer customer lifetime experiences?
Huiming Qu: Absolutely. We certainly care about and wanted to really improve the search recommendation relevancy and, every time people landed on our site, to provide a better experience.
These projects — sometimes it takes multiple sessions, multiple days, multiple weeks. So it certainly is a customer journey. We do want to remember as much as possible where the customer stopped: What is in the cart, what are some of the prior searches, and what are some of the prior visits? And did this customer actually click on some of the emails that we sent or, outside of Home Depot engagement, other websites about some of the marketing messages? It’s a holistic experience: The more we understand about holistically where the customer is in their journey, the better we can serve them.
Sam Ransbotham: Getting this holistic understanding seems complicated. What infrastructure and project management does getting that understanding require?
Huiming Qu: It takes a big effort of engineering to change some part of our infrastructure to accommodate the new algorithms that we’re delivering. The algorithms are already being developed offline. We know this is also something we want to embed into our search experience to serve it to our customer. But initially, when we design this, it takes engineering an estimated six to 12 months to deliver that change, because it’s a huge change.
The algorithm iteration also takes months to be developed. But when we look at the road map, when we actually stack them up, [it] will take years to get there. So how we get there within, actually, the next 12 or 18 months is the challenge — to really see, as early as possible, the test [version]. What is the extremely, extremely light test version that we can get to? To get there much faster, we need to ask ourselves, “What is this really light version?” Maybe we can deliver [it] in 30 days, but we can actually allow the data center to see that infrastructure change is not for the final deployment of that architecture to really host the algorithm. It is really just for the testing, so that we can actually move a lot faster. And it takes iterations for us to get [it]. … It takes time to test.
These are actually customer experiences that are interweaving with data scientists’ work, user experience work, product and engineering, to understand if it’s even possible. It takes a cross-functional team iteratively to move a lot faster — to break down that bigger problem, bigger goals, to many smaller ones that we can achieve very quickly.
Sam Ransbotham: You mentioned cross-functional teams. I was kind of curious. … You also mentioned customers a lot, but what effects do you see within your organization about places … you’ve adopted artificial intelligence in some areas; what effects does that have on the organization itself? Are people worried? Are people happy? Are they more siloed? Are they less siloed? Do they have terrible morale or greater morale? What’s it doing to the organization when you put these things in place?
Huiming Qu: I think in general, truly, everyone believes in the power of machine learning and AI. We have this … drive to have the human in the loop because domain knowledge in home improvements is extremely important. This process is really helping the data to be prepared for really unleashing the power of data science and algorithms.
Shervin Khodabandeh: So the algorithm can help, but the algorithm on its own is also quite powerless, or it will take a long time of unnecessary learning. This is really fascinating, what you guys are doing with all of that. I have to say I have a bunch of tools, but I’m sure the algorithms would be able to tell me, “Hey, you actually have a different problem, and you’re using the wrong tools,” and this is why, Sam, my projects are never-ending. Another reason the projects are not ending is because you don’t know what you’re doing.
Sam Ransbotham: Blame it on the tools: “It’s all the tools’ fault; it’s not mine.”
Huiming Qu: Yeah. We had that. So some of the customers — it could be myself; I don’t know what I’m getting myself into. I install the mirror and the curtain rod in the wrong way, and it would just fall. It takes probably iterations to get it right. But I think [for] any Home Depot employee, we feel the responsibility to really deliver that right product to you with basically an understanding of “This is the right project that you can do DIY, or you may want to consider services provided to you.”
Shervin Khodabandeh: I have a pseudo-technical question for you: As organizations think about investing in these kinds of capabilities and building the right teams around it, do you feel it’s more a data science investment and hiring of data scientists and data science capabilities, or have the algorithms been commoditized enough that it becomes more about connecting and building the right APIs between data to algorithms, to production system — hence, it’s sort of a different profile of people, much more engineering oriented than, I would say, Ph.D.s in data science? Do you have a point of view on that?
Huiming Qu: I think probably today the titles may not be indicative in terms of the skill sets needed for the title, so I’m not going to be strictly using the title. But I would say in the future, we would probably need more of the interdisciplinary kind of role — people who can connect the dots — because we don’t necessarily want to always have the Ph.D. scientists create customized solutions for many of the problems that we’re solving. As machine learning algorithms are getting more commoditized and accessible, that is not the key to the success.
The key to success is, how do you feed the data? What type of data did you feed [to it]? How do we not do it repetitively and have a lot of redundancy, right? You do it in parallel to solve similar problems. The way we see it is, how do we really host that solution, feed [it] with the right data, build it once and use it many times, in a scalable way?
We also understand a data science platform will help us to have the data scientists focus on the right problems and the right step of the problem. That step may not be creating a new algorithm. That step could be helping our product team, helping our business, to understand what is the right problem. What is possible? There’s a slight role change in whichever title they’re in, but collectively, we’re going to build one solution to help serve many of the machine learning applications and help us be smarter … whether it’s assortment choices, pricing choices. It could be product collection selection choices and providing [them] to search results in a very scalable way. I don’t know whether I exactly answered your question.
Shervin Khodabandeh: I think what I’m hearing is, you’re saying it’s going to evolve to be much more about connecting the dots and really stepping back and saying, “What does it take to get value, whether it’s for merchandising or pricing or marketing, and hence, what do I need to connect to what?” much more so than, “How do I improve the algorithmic efficiency of this model by 1%?” It’s much more about connection.
Huiming Qu: Sometimes the business area to improve that model with 1% is needed. In the area of search, it’s needed. Because it’s a very deep problem, it has very specialized techniques to solve it. It takes a combination of search algorithms and search platform integration to solve it. These are the areas that we need to go deep, but [for] many of the problems, we’ll see a tremendous value to go wider and consolidate to have better solutions in integration.
For Home Depot, we have over two-thirds of [our] revenue influenced by search. The traffic and the amount of impact it will make … that can actually reflect that 1% to be very impactful. [In] other areas, you may not need this — kind of going very deep and having a very specialized team to do it.
Shervin Khodabandeh: Yeah, thanks for correcting me on that, actually. I think it’s very well said. Integration is one axis and then depth is another.
Sam Ransbotham: Huiming, you clearly have a great background to contribute to both the integration and depth axes of AI problems. Can you tell us a bit about that background? What brought you to The Home Depot?
Huiming Qu: I think my career is probably a series of happy accidents. Starting from 19 years ago, I came to the U.S. for grad school. In my last year of my Ph.D., I took this data mining class — that’s from professor Christos Faloutsos. That data mining class started my journey in data mining and machine learning. Since then, I’ve worked on a wide variety of interesting problems from supercomputer resource optimization — these are enterprise social network analysis and pricing problems, marketing problems, and, today, some of the very challenging e-commerce problems that we are solving.
My first job, [at] IBM, I was in the service department — part of [the] service department; back then, IBM [had] 50% of their business around service. After that, I had the opportunity to work in a startup, which is very different. I remember when I interviewed, I was asked, “Why do you consider [moving] from IBM to a small startup of, like, 50 people?” I really wanted to experience that culture, and the type of problem is very different. I learned so much building things in production, which I still think is a very rewarding experience today.
Then I worked in [the] financial [services] industry as well, which is actually another accident. I didn’t plan to go into the financial industry, but I guess it’s hard to avoid because I was in in New York. It’s a very good experience in the middle between IBM research and a startup — very fast paced, with a great culture, solving very interesting problems, having access to a huge number of problems embedded but a huge set of data as well to understand the customer’s interest and really serve them the best in a serendipitous way. But Home Depot, I think, is going to be the closest to where I really love, because it’s influencing consumer behavior. It’s really impacting hundreds of millions of people’s everyday lives. I dreamed about this, and it’s truly making those impacts that’s really powerful.
Shervin Khodabandeh: That for sure sounds very rewarding and eventful. Huiming, I want to take us back to comments you made about transitioning from research to a fast-paced environment and putting things in production. Can you comment a bit about what that is like for folks that are moving from a highly academic, very research-oriented — like pushing the boundaries and the frontier — to an environment where you’re trying to work in a context of a company starting point and challenges and getting to value. What are some of the lessons learned there that you think [are] helpful to keep in mind?
Huiming Qu: I think the most important thing is, the reward system changed. When you’re a researcher, it’s very rewarding, but it’s a different kind of metric. You’re rewarded by working with extremely talented people. These are people who later on go to great companies, working on very challenging projects and contributing to society. You’re rewarded by the number of publications and patents. I still have my patent plaque from IBM — not on the wall! — but it’s a very rewarding experience from that rewarding system.
When I joined a startup, I think I didn’t appreciate it, but the [reward] system is also very rewarding; that is, actually pushing the code into production and seeing the problem that I was specifically working on — this real-time bidding for these ads. It’s very fast paced to deliver the code, to deliver new algorithms, and to test and to iterate very fast. It’s not going to be measured by papers or publications. And even, actually, the people you’re surrounded [by], how they’re recognized — [they’re] rewarded by, really, the results and how quickly we can ship and how quickly we can react to some of the new requirements provided by our customers. I think that was really a shift. You need to quickly shift how you’re looking at those experiences and really react to those [reward] systems.
Shervin Khodabandeh: It’s actually quite interesting, because the whole domain of what we’re talking about — AI and machine learning — is about feedback. It’s algorithmic feedback, but you’re also saying process feedback and user feedback and market reaction.
Huiming Qu: I think that the team success is very, very important. Throughout the years, we have tried to figure out, how do we really create an environment that helps create innovation — create innovation driven by machine learning and data science — to be successful in that organization? That has been one of the key metrics in the reward system, because we know it’s going to be helping us to be even more successful in the ultimate goal we are driving [toward].
Ultimately, it’s going to be well received by our customer because … having them happy is the ultimate success for the business, too. But one more thing [I would add] is to have the team that is successful, to make it happen. … Probably [my biggest] contribution is, I hope that I keep them happy and excited about the problems they are solving and hopefully creating that environment that helps those scientific dreams to be realized. … The [reward] is seeing [the] happy faces of the customer.
Sam Ransbotham: Huiming, many thanks. You mentioned both the depth and integration, and clearly Home Depot is fortunate to have someone like you that excels at both the depth and the integration part. We really enjoyed talking with you and learning about your background and learning about Home Depot today. Thanks for taking the time to talk with us.
Huiming Qu: Thank you for inviting me. It’s [been] a very nice conversation.
Shervin Khodabandeh: So that was quite an insightful conversation, Sam. What did you think? What resonated with you very strongly?
Sam Ransbotham: Her enthusiasm for the problem really showed, and this idea of the reward structures and how [for] every problem that she worked on, even in different organizations and in different areas, she had to be very careful about the tying the incentives and the reward structures to the problems at hand. I was surprised at the variety of different feedback loops that that created.
Shervin Khodabandeh: I think that’s also the first thing that stood out for me. And also, you’ve got to be so much more focused on getting it out to production, not letting the perfect be the enemy of the good, and also incentivizing your teams — who also, at some point, came from an academic background — to sort of adjust to this new reward mechanism.
Sam Ransbotham: I know you like to call me academic, Shervin, but I could see the appeal of getting some fast feedback on things. We work on projects — even the projects that you and I work on — they take a full year from the time we go from surveys to interviews through a research product. She’s talking about changing things and getting feedback immediately — I’m jealous. That was pretty beautiful.
Shervin Khodabandeh: And actually, when you talk to data scientists, that’s really what they love about their jobs, the ones that are focused [on] building solutions that go into production. And then I also liked that she was quite cognizant of the importance of … it’s not just about getting some solution to production. I mean, it has to be good and it has to be efficacious, and there are times where the algorithmic efficacy is really, really important. You’ve got to be able to understand which situation you’re in, which also, to me, opens up a dialogue about the skill sets and the attributes of a successful AI specialist or data scientist — that not only do they need to have very solid technical skills and pedigree and background and all that.
Sam Ransbotham: And you’re also foreshadowing, because we know that PepsiCo is coming up [in the] next episode, and the variety of talents needed to solve problems also comes out very strongly in that discussion.
Shervin Khodabandeh: That was a good segue.
Sam Ransbotham: Today’s discussion of talent needs segues well into our next discussion with Colin Lenaghan from PepsiCo, as he talks about the technical, organizational, and commercial talent PepsiCo needs to accomplish its objectives. Please join us.
Allison Ryder: Thanks for listening to Me, Myself, and AI. If you’re enjoying the show, take a minute to write us a review. If you send us a screenshot, we’ll send you a collection of MIT SMR’s best articles on artificial intelligence, free for a limited time. Send your review screenshot to firstname.lastname@example.org.