Home Depot Finds DIY Success With Vector Search


Like all major companies, Home Depot has a list of IT projects it wants to tackle. When the COVID pandemic struck two years ago and e-commerce activity grew, it accelerated one of them in particular: the development of vector search algorithms to extend keyword-based searches on its website and mobile app. Since going live, its proprietary vector search engine has delivered exceptional results and, most importantly, a more relevant search experience for visitors.

If you’ve ever tried to search for a specific part or obscure product on a website, you know how difficult it can be to find it. Unless you’re on the exact same page as the company in terms of the words it uses to name and describe its products, you’re unlikely to find it on the first try.

The usual response to this common obstacle is to try different words, which is popularly known as “Google-fu”. If you are good with words and persistent, this is usually enough. But if you’re busy or just can’t think of new words for whatever reason, you may end the search before you find what you’re looking for.

Huiming Qu, vice president of data science and analytics at Home Depot, is familiar with this phenomenon. “It’s very hard to describe things,” Qu . says Datami† “A lot of the products are so specialized.”

With over 2 million products, Home Depot has more than its share of obscure items, and it sometimes struggles to help people find them. Whether it’s a slant-down rod for a Hampton Bay fan or an 18-volt lithium-ion battery for a Milwaukee cordless drill, Home Depot’s search engine has its work cut out for it.

Home Depot sells more than 2 million items online, with each store stocking about 35,000 items (Mihai_Andritoiu/Shutterstock)

In addition to the sheer amount of random widgets and number of searches, Home Depot is challenged by the diversity of its users. Professional contractors use different words than the do-it-yourselfer on weekends, and salespeople also describe products differently than Home Depot itself, Qu says. There are also geographic differences in how people talk. And did we mention spelling mistakes?

“We have this contest of what are the most misspelled words?” says Q. “[For a word] as simple as ‘window’, there could be 20 ways to spell window.”

Enter the vector search

Until recently, Home Depot’s IT professionals loaded as many of these creative product descriptions, regional variations and misspelled words as possible into their search engine index, crossing their fingers to help people find the right product. This can be considered a brute force approach.

But over the past few years, a more elegant approach has begun to take hold. Dubbed vector search or sometimes neural search (why stop at one sentence to describe things?), this new technology uses a fundamentally different technique to match users with the items or products they are looking for.

Rather than boosting the search by attempting a direct one-to-one match of keywords, a vector search engine tries to match the input term with a vector, which is a set of attributes generated from objects in the catalog. In this regard, vector search leverages the predictive power of deep learning and a large set of sample data to better understand what a user is looking for.

The advantages of vector searches come from the fact that each vector can have tens to hundreds of dimensions, each describing some aspect of an item in the catalog. While this technique requires the ability to work with big data and requires more computing power, the net result is the delivery of search results that reflect more nuance and context in the search space where words and users meet.

Vector search uses neural network techniques to improve the relevance of search results (Evannovostro/Shutterstock)

Searching Vector has been on Home Depot’s to-do list for a while now. But the wave of online business from COVID provided the perfect opportunity to take advantage of this new technology, and in 2021 it rolled out the first version of its vector search engine, Qu says.

Early results are promising, Qu says, especially when it comes to detecting user intent and helping find hard-to-describe or obscure parts or products.

“When you have a four-word search term with all that complexity, it’s very hard to find exactly with keyword search,” Qu says. “This just screams understanding intent. What does that four-word search term mean? So that’s where vector search comes in. It’s not just about literally understanding the words.”

Fanning Out Vectors

With a more intelligent vector search engine that extends the brute force of keyword search, Home Depot’s Intent Search engine has a better chance of presenting the right product to the customer in the very short time it has to work with.

Take, for example, the classic Home Depot use case: the installation of an outdoor ceiling fan. Home Depot has many different types of fans. The question is, which one should he show to the customer?

With terabytes of historical data to work with, Qu’s vector search engine can uncover hidden links between products, such as sloped ceilings, ceiling fans and downrods. So when a potential customer who needs to match their ceiling fan purchase with a downrod of a certain type and length does their search, the engine will return more relevant results.

Even the simple ceiling fan downrod can quickly get you caught up in a quagmire of search engine complexity (Image source: Home Depot)

The vector search can also provide other information about the problem, including past searches, Qu says. Perhaps Home Depot knows a particular customer is in the midst of a patio renovation, immediately narrowing the search to outdoor ceiling fans. And if at any point a pitched roof was mentioned in a search, the vector search engine manages to prioritize products related to it, rather than fans designed to be installed against flat ceilings.

“I would call it a combination of putting together the history of what we know about a customer and then connecting that with the product knowledge,” Qu says. “We really took the friction out of asking the customer to specify, ‘I specifically need this 1.5m downrod ceiling fan.’”

The results were significant, according to Qu, who shared some specific KPIs Home Depot uses to track its search results. For example, after implementing the vector search engine to power its ‘Intent Search’ service, the company has seen a 13% increase in nDCG, or cumulative profit at a discount, which is a measure of ranking quality. It saw an 8% decrease in search reformulation, which is a measure of search friction, and a 45% decrease in the proportion of complaints related to the relevance of search results. Engagement with the best search results has increased.

“We’ve seen a huge improvement in our search relevance,” Qu says. “There may be customers [eventually] finding their results after three searches. Now they can only do it once.”

Your own knowledge base

The rollout of vector search algorithms was not easy and required a significant effort from Qu and her team to develop the semantic machine learning model, which is built in Python and runs on Google Cloud.

“It’s definitely one of the most challenging projects we’ve worked on because it’s not… just implementing an algorithm, it’s the platform change,” she says. “It changes the way we index. It changes the way the data pipeline has been. So it’s very systematic changes, a collaboration between data scientists, our machine learning engineers, our search engine engineers, to really roll out this project.”

Huiming Qu is the vice president of data science and analytics at The Home Depot

Improving the relevance of search results was the first goal, followed by improved personalization. But more projects are in the works, each requiring multiple teams to work together, she says.

For example, Home Depot is also working on applying computer vision to search results. The benefit of being able to detect patterns in images may not be immediately apparent with a word-driven search engine. But as Qu describes it, it’s all about grouping similar items together, and that similarity can be centered around a specific visual style.

“When you buy chandeliers, it’s hard to describe what kind of chandeliers you’re looking for,” Qu says. “You’d say, ‘I’ll know when I look at it.’ And when you land on one, there are traits, and we can basically recommend similar people.”

Sometimes the hints are more direct. For example, if you specifically search for “Mid-Century Modern,” Home Depot will tell you that you’re only interested in products with that tag. That doesn’t necessarily make Qu’s job any easier, though, as her team still needs to do the work of labeling all items with the correct style (tagging all those items manually would probably be too expensive).

“We are improving product features using computer vision. We can create those text attributes that are translated from the image attributes,” she says. “Vector search is a big asset for us, but behind the scenes we also have a lot of recommendation algorithms, which understand what the accessories and collections of these other products are.”

These new technologies, such as vector searches and computer vision, are no substitute for traditional keyword searches. Home Depot uses an ensemble of different search technologies that come in handy when needed. There is no way of knowing if your specific search was powered by one search engine or the other. It all comes together at Home Depot under the banner of its knowledge base, Qu says.

“A lot of these technologies are all hosted together,” she says. “This is really our home improvement knowledge base, and in return it will help improve search, also part of the visualization experience we provide to our customers, and improve recommendation. Many of these are product discovery capabilities that we provide to our customers. ”

Related Items:

Are neural nets the next big thing in the quest?

Elastic gets new vector search and NLP capabilities

Microsoft applies Deep Learning to Vector Search

Leave a Reply