Thursday, October 19, 2017

From Physics to properties: Making the move to a UK rental start-up from CERN

by Boris Mangano, Senior Data Scientist at Movebubble

When growing up, there were two dreams of mine; to become a physicist and to work in a start-up. Having spent over a decade in the traditional world of physics, working for CERN, the international particle physics laboratory, on the team that uncovered the Higgs Boson it was time to realise the second dream.

Whilst from the outside the move from physics lab to UK rental start-up may have seemed an odd one, it offered the chance to implement many of the skills and data analytic techniques I used as a scientist in a company where I could make some real impact.

The start-up I chose was Movebubble, the only renter-dedicated platform that enables users to move homes in just a few taps.

As an online application used by renters and agents alike, Movebubble had the scope to not only gather large amounts of data but also to utilise and analyse its data to help improve users’ overall rental experience for good. Renters can currently discover thousands of properties in real-time, book viewings and secure a home hassle free.

I saw a huge opportunity for how data analytics and machine learning could not only enhance what Movebubble was currently doing but improve its service to renters.

Machine learning bursting the rental bubble

Machine learning (ML) has become something of a buzzword over the past few years. In general, ML is the part of computer science that focuses on software able to learn without being explicitly programmed.

The two main applications are ‘classification’ and ‘regression’ and we currently use both forms to help renters.

Using ML for classification, we can predict genuine property listings from common ‘bait and switch’ property listings. During lulls in the market, it can be tempting for agents to leave unavailable properties on online markets to ‘bait’ renters and then ‘switch’ to offer them alternative properties.

With ML, we can classify which properties online are likely to be of this type and save renters from wasting their time by removing them from the platform.

Utilising ML for regression models enables you to predict quantities from a set of inputted data.

Regression application of ML at Movebubble could enable us to predict the ‘fair’ rent of a property given its location, the number of rooms or type of furniture in the property.

This will be of particular help to users who are perhaps moving to a new city where they don’t know much about the area, or for users who have a particular set of property criteria.

Whilst you could write classification and regression programs explicitly, using a long sequence of if-else-then statements, utilising machine learning is far more efficient.

The programmer acts as a supervisor, giving the machine a series of examples and the corresponding correct answers effectively ‘training’ the machine until it learns how to make generalizations.

Once it has learned for example, that average rental prices in Clapham for a three bed flat are £1,500 a month, it is able to make predictions about previously unseen data, like new similar properties recently added to the market.

In order to improve the users’ experience of renting in the UK, my job is to first harvest from the platform the data that contains the most valuable information about the problems facing renters today.

After writing the initial algorithms, monitoring that the machine is learning is essential, as it understanding the pace at which the machine is learning.

Does it require additional data to improve its performance or increase the speed it is learning at? Has it already reached the point of diminishing returns? Often the answer is not more data, but different type of data.

Also important to get right is the type of machine learning algorithm that best fits the problem you are trying to solve.

There are a long series of different algorithms, random forests, support vector machines, logistic regression, shallow artificial neural networks etc. that have been used routinely for at least 20 years now in different industries. For the business tasks faced by Movebubble currently, conventional ML algorithms are more than enough for the time being.

I believe more advanced machine learning as deep learning should only be used for specific tasks that really require the ‘fire power’. Otherwise the overhead costs associated with the implementation and the training of the algorithm doesn’t provide sufficient ROI for its results.


Click ‘Next Page’ to read about renting powered by artificial intelligence.