Houzz Leverage Deep Learning to Level Up its Game in Online Marketplace
Over the past decade, machine learning has profoundly disrupted online ecommerce shopping experience. It enables ecommerce businesses to create a more personalized customer experience by constantly training and refining their consumer datasets, leading to better conversion rates overall. However, for a business model like Houzz, it faced some unique challenges.
Started through word of mouth, Houzz is an online marketplace platform that connects professional interior and exterior designers with the house owners by charging the 15% commission fees. Over the years, it has also adopted the pinterest model to provide inspirations for home design. Very soon, the company realized that this commission based model was not sustainable.“People come to Houzz because they want to get everything they need to improve their homes in one place, from inspiration to execution,” said Alon Cohen, Houzz co-founder and president [1].
October 2014, the company announced the beta launch of its newest revenue channel, the Houzz Marketplace, which will allow users to directly buy some of the products they see when they browse the site. The service currently features over one million products from a large variety of sellers [2].
However, with over 15,000 merchants selling on Houzz today [3], consumers could be struggled to find the pieces that satisfy their decorating needs, in particular those who are looking for furnitures that match unique design styles or material qualifications.
With marketplace being the major revenue driver of the business, how can Houzz drive better furniture shopping experience in a marketplace with such a massive amount of product offerings. How can they nudge consumer behavior from inspirations to conversion in a smarter way?
For Houzz, deep learning is the answer.
How did Houzz do it?
September 2016, House introduced Visual Match on Houzz.com and in the Houzz app [4]. Visual Match scans photos to identify similar products — from tables and sofas to mirrors and plumbing fixtures — and shows you examples of those products that are available in the Houzz Shop [5]. With Visual Match, consumers can easily discover and buy various types of products and materials that inspire them in photos. See the featured image below [6], a consumer browses this particular image for decorating inspiration. Simply hovering over the lamp in this photo, this consumer can click the magnifying glass if he is interested in the product. Based on the click response, the web server will then quickly make a call to its product image database.
By applying its proprietary imaging scanning and matching algorithm (based on metadata collected from all the lamps in its product database), it quickly displays eight different lamps with a very close visual match to the lamp in the original photobook. Consumers then have the choices to click the image of lamp, which then get directed to to ecommerce section of the site. They can browse through the product details and purchase this product from there.
While Visual match has made it so much easier for consumers to discover and buy products on Houzz in the short run, the company is also thinking of further revolutionizing the online furniture shopping experience in the long run. Houzz recently started to experiment with augmented reality that leverages a catalog of 500,000 images for people to view, move and install in their virtual rooms [7]. The product images will appear more “lifelike” than ever in the virtual rooms, ultimately driving higher purchasing intent and more satisfying furniture shopping experience.
A path forward
While deep learning has created significantly more positive browsing and purchasing experience for its consumers, there are still problems remain unsolved. Image below [8] is a perfect example.
When a consumer is interested in the coat hanger behind the sofa, he hovers to the object and clicks the image tag. Unfortunately, only a list of lamps are displayed. One assumption could be that the product database doesn’t have enough metadata info stored for coat hangers; or it could be that the product offerings themselves are very limited. One recommendation to improve the visual match quality is to collect multidimensional meta tags for the products and assign priorities to different training rules in its matching algorithm. In addition to just using visual proximity as the key matching factor, the database could also group all the products into different product-SKU categories. These product categories could have higher priority than that of visual proximity when image matching algorithm is applied. In the long run, Houzz could also develop functionalities that enable user-generated filter rules based on product style, price range, materials etc to further improve image matching quality.
The key open questions then are: 1) How cooperative will designers be when they find out that all the photobooks they provide have now become the marketing tool for Houzz while their original purposes were to attract commissions from consumers; would this affect the product metadata collection process? 2) How much authority would consumers be willing to give up when browsing product offerings? Should product offerings be displayed based on user generated rules rather than purely algorithm based rules that consumers have no control of?
(Word count: 793)
References
[1]“Houzz Raises $165M Series D Funding Round Led By Sequoia To Fuel International Expansion.” Techcrunch, September 14, 2016.
https://techcrunch.com/2016/09/14/houzz-now-uses-deep-learning-to-help-you-find-and-buy-products-in-its-photos, accessed September 2016
[2] “Houzz Raises $165M Series D Funding Round Led By Sequoia To Fuel International Expansion.” Techcrunch, September 1, 2014. https://techcrunch.com/2014/10/01/houzz-raises-165m-series-d-led-by-sequoia-to-fuel-international-expansion, accessed September 2014.
[3] “Find furniture that suits your style with Houzz’s new visual match tool.” Digital Trends, September 14, 2016. https://www.digitaltrends.com/home/houzz-visual-match, accessed September 2016.
[4] “Houzz Introduces Visual Match, Making it Even Easier to Discover and Buy Products on Houzz”. Houzz, September 14, 2016. https://blog.houzz.com/houzz-introduces-visual-match-making-it-even, accessed September 2016.
[5] “Find furniture that suits your style with Houzz’s new visual match tool.” Digital Trends, September 14, 2016. https://www.digitaltrends.com/home/houzz-visual-match, accessed September 2016.
[6] Image source: https://www.houzz.com/photo/131141930-seminary-road-traditional-kitchen-denver
[7] “A new ARKit app from Houzz brings 500,000 objects to moveable life”, Techcrunch, September 19, 2017. https://techcrunch.com/2017/09/19/a-new-arkit-app-from-houzz-brings-500000-objects-to-moveable-life, accessed September 2017.
[8] Image source: https://www.houzz.com/photo/21475570-sliding-door-sliding-panels-woven-woods-farmhouse-living-room-oklahoma-city
I found this article interesting, in that it asks us a question about how much control AI should have in order to make any decisions related to aesthetics. Utilizing huge database, machine learning could recognize patterns of interior design and make suggestions about a better space. However, as you pointed out, how much control users are willing to give up depends on how precisely the AI could address their preferences. I also agree with the point that the service casts doubt on how corporative designers will be while they know their ideas are being consumed to generate revenues for Houzz. I believe that creating the ecosystem that delivers benefits for both customers and designers is a key issue here to integrate the visual service with the company’s core business.
I think it is very interesting that they recommend furniture based on visual matches. I wonder if they can extend their deep learning algorithms to take into account designed feedback. For example, designers could look at the images of the rooms and make recommendations on what furniture they think would fit the setting. The algorithm can then take this data and try to “learn” how to design a room just like a designer would. The issue here might be the same one that you pointed out: how to get designers onboard?
I appreciated this article and the implications of machine learning in eCommerce. It appears that at this point, Houzz is primarily using machine learning to recommend additional or similar items to people based on what they have already viewed. This technology could definitely be taken further past just product aesthetics and to price preferences, items that are purchased in tandem with previous purchases, etc.
However, I do wonder if it might be more difficult than one might realize to understand consumer preferences when it comes to aesthetics and design. Speaking simply for myself, my tastes are extremely varied and I don’t know how effective a website would be at understanding all my preferences based on just a few items that I looked at or bought.
I also wonder how much buy-in you need from the designers on your website. Without those designers, you don’t really have a business since your website is predicated on providing designer items. Would this type of machine learning that sends potential customers to competitors anger designers? The mix of how many designers you have would almost certainly changed, which may not be terrible for Houzz, but is something to consider.
I loved this article. I believe if Houzz can pull this off, it will have found a unique platform for home buyers / sellers to generate revenue in between home sales (which typically have long sales cycles). As future home buyers review homes, they may be spurred into a few more impulse purchases before even getting a mortgage and closing on a home!
What the company has to do first is validate the concept, which is still in progress, as pointed out by the author. But not much further than whatever it takes to work (e.g. a “minimum viable product”). By fixing key issues, the company can have a sellable product to the very people who post those photos. Then encourage home sellers to do any tedious manual tagging work for Houzz, expecting the promise of commission for any items sold. Furthermore, Houzz can encourage “look-a-like” on-demand manufacturers to work on the other side Houzz’s marketplace, even supporting them with consumer data on what new home buyers are focused on when looking to decorate a home.
I am excited to learn more about this project as it develops.
Great article. Regarding your open question 1, I believe that this can potentially lead to future concerns over the Intellectual Rights for designs as this model starts eating away from the role of the interior designer.
On the one hand, as this AI technology progresses, it is conceivable that AI could be able to replace the designers all together – which the designers should try to avoid.