Meet the Man behind Housing.com’s Recommendation System

Housing.com claims to be the country’s largest full-stack digital platform that supports consumers in their home-buying journey. Founded in 2012, it’s one of the largest property platforms in India. To find out more about how its recommendation system works, what challenges they face and how generative AI can help property platforms like Housing.com, Analytics India Magazine reached out to Sangeet Aggarwal, head of products & design, Housing.com. 

Aggarwal has over ten years of experience in product and technology, and currently leads the product roadmap for monetization, ad tech, customer experience and engagement at Housing.com. 

AIM: Can you please tell us about the recommendation system of Housing.com?

Sangeet Aggarwal: For users, the interface still stays the same, but the recommendations have become much smarter than before. If I were to talk about two or three years earlier, our recommendation systems were extremely simple. The user would just look at a budget and locality and we would find the matching listings within close proximity, with which there was a good chance that multiple users would see the same recommendation. 

Before we actually had moved to a more heuristic-based model, where essentially, the idea was to try to make some parameters of what the user was searching but not essentially an ML algorithm to back it up. This calendar year, however, we have now moved to a full-fledged ML engine. So it takes into account the thousands of signals, in terms of what other users have done. About 50 to 60 different data points of the other user’s journey and then tries to give you a recommendation. 

AIM: What are the challenges you’ve faced while building the recommendation system?

Sangeet Aggarwal: In terms of challenges, from a user perspective, there are some challenges both internally and externally. For instance, one challenge which poses a problem for any ML recommendation engineer builder is: how do you personalize the UI on the first visit? Once the user has visited the site multiple times, you know something about the user and the recommendations become more personalized but for the first visit, you pretty much have to rely on what other users may have done from the region or some other parameters.

Another challenge I would say is what happens offline is something that we have no idea about. For instance, if someone looks up a home on our webpage, and then decides to go and check it out, we have no idea what happens there. The whole perspective changes, the goals change and when you come again to our website, we are still relying on the data from your last visit; but a lot has changed in the meantime. And thus, the user might not prefer the recommended products which it did earlier. I would say, these are two big challenges on the consumer front.

AIM: Housing.com claims to have the image auto-auditor and still there are plenty of duplicate images and listings on the site. 

Sangeet Aggarwal: I would sort of position this as a journey. With any new innovation in the market, it takes two-three years to get it to the level where we start seeing results. Given that real estate is one of the largest industries, the potential for clickbait is always there. What we have recently launched is helping us crack this particular problem on the basis of internal data. We’re trying to identify duplicate listings, fake listings, clickbaits etc with our in-house developed image auto-auditor that is created on an auto-rejection and correction pipeline for various image use cases to achieve over 90% automation in the auto-audit process for flat images.

The problem has certainly improved, but yes, it hasn't been zero. We can’t claim it’s 100% working out there, but it certainly moved in a positive direction for us. We’re planning to invest in the particular area and given the size of the problem, it’ll take at least a couple of years to tackle the problem. 

AIM: What is your opinion on generative AI? 

Sangeet Aggarwal: Every week, you wake up and something new comes up in the market. It's really amazing seeing the exponential growth in the AI industry. So, we are exploring that. I’m pretty sure a user today wants the conversational AI sort of interface, and once this becomes more open source I think we'll be able to leverage on this trend.

AIM: Can it be used in the tech property market?  

Sangeet Aggarwal: For the people who read all the property details, it’s becoming a long page for them on the platform. As the attention span is reduced and users are more inclined towards the 30 second videos, we were wondering if we can have an engine where it takes a lot of inputs from content around property, descriptions, pictures etc, and then builds a quick video around that.

But again, each property is unique and has a unique flavor. We’re really trying to figure out what can be done and it's a long journey ahead, but again, this can be a possible use case of generative AI in the online property engines.