SUSS talk by Setu Chokshi, Head of Data Science at PropertyGuru Group

Listened to Setu Chokshi, the Head of Data Science at PropertyGuru Group at a talk organized at SUSS on AI and IoT by IoTSG, an IoT & Advanced technologies focused Special Interest Group. At any given time PropertyGuru has about 60000 live rental listings and Chokshi gave a brief overview of what’s happening inside PropertyGuru esp wrt AI. 

Most of the action in the property business as it is structured today, happens in the discovery phase. This key step has traditionally been pivoted around the location factor and PropertyGuru is attempting to use AI to enhance the end buyer experience around this. He requested me to remove some of the details as those features were at a beta stage. An existing PropertyGuru feature that I really liked on the platform is the historical actual transaction values for each apartment complex in Singapore. Something like this helps all stakeholders make more informed choices and bring stability to the overall market. Compared this in India, there is a stalemate like situation where Buyers find it difficult to buy a property at their target price, and once they consummate the deal, they find it equally challenging to find a buyer for the property. Setu talked about other life features like area finder which he claimed was a useful underutilized feature on their website. On a closer look,  I thought a few things like making the capability available on the phone app and using the insights on the social platforms could increase the utilization of this rather powerful feature.

They also do a lot of consumer analytics aimed at measuring the effectiveness of the Ad campaigns being run on behalf of the agents. A key effort is around identifying the stage where each customer is in the overall purchase lifecycle. The related Consumer Lifetime Value is an interesting idea to better segment and target the consumers with the right information at the right time. All that data is aggregated into a format targeted for the developers.  

Chokshi then described an interesting Beta project which has very limited direct commercial benefits for the company. PropertyGuru Lens is an Augmented reality setup that allows users to identify the condominiums by snapping an image of it. A typical image detection model based on computer vision is 100s of MBs in size and extremely resource intensive. So the challenge was to develop a tool that didn't take up too many resources and heat-up the consumer mobile phones. They optimized the solution by using the geo-location of the device. However, there were serious issues wrt to privacy regulations and the latency errors of the device compass. Despite these constraints, they optimized the solution with correction algorithms to create a 2MB on-device model that had a latency of merely 6ms on an old iPhone 7. The performance would apparently be better on the newer iPhones.

Now in Singapore, each building has a unique postal code and if you know the postal code then you can easily find the information of PropertyGuru listing easily. And with commercial GPS solutions, it is very easy to find the postal codes of any building in Singapore. Given that using a potentially error-prone Augmented reality solution adds little additional commercial benefit. But the experience of developing this tool helped with the capability development of the PropertyGuru AI team.

Eventually these skills enabled the team to create a faster and cheaper means to train models using a subset of the ImageNet visual database with 1.4 million image and some 1000 categories. Typically training the models using this database took a few weeks if not months. In addition to these tweaks they also switched from an internal Server to an AWS lambda serverless architecture. The direct cost reduction involved was from $3-4000/ month to low three-digit figures. However, the bigger advantage was that the AWS solution was more scalable, repeatable and possibly had shorter deployment times. However, as a word of caution, Chokshi did highlight that there were some technical issues with this AWS lambda setup.

There other interesting comments, like how Chokshi’s team worked on preventing racism on the platform. While admitting that it was a cultural issue, he highlighted how his team was working on preventing any racist ad listings on their platform. Additional work in this area has even caught the attention of Singapore’s policy makers.

In addition Chokshi mentioned other research areas of interest for the company:

§  Prevention of inappropriate language, agent phone numbers and competitor information on the platform.

§  Image caption technology to make the listings more meaningful.

§  Applied reinforcement learning

§  Fraud Listings: He didn’t elaborate, but I am assuming that it revolves around very attractive non-existent properties by Agents to attract buyers. 

From the reaction of the audience, I got a sense that there was some good technical work. But instead of basking under this glory, Chokshi humbly admitted that they were standing on the shoulders of other giants who had done some foundational work which made these results possible especially with respect to the advances made in the recommendation engines and the property discovery processes. His humility was on display consistently throughout the talk and it rather was endearing.

The final comment I will make is that there is a tectonic shift in the attitudes towards property purchases from Capex to Opex. Youngsters today prefer renting to owning and this shift means that the opportunities will move from buying and selling to renting and leasing and the allied services around that. That will operationally more challenging but more consistent revenue streams involving smaller value transactions. Companies with people like Chokshi will be in a better position to leverage these changes and opportunities.