eHarmony: How machine learning is leading to better and longer-lasting love matches

Machine learning is being increasingly employed to help consumers find a better love match

eHarmony creates more personalized matches
Relationship-minded online dating site eHarmony recently upgraded its cloud environment to use CDH and the Intel Xeon processor E5 family to analyze a massive volume and variety of data. The technology is helping eHarmony deliver new matches to millions of people every day, and the new cloud environment accommodates more complex analyses to create more personalized results and improve the chances of relationship success.
eHarmony creates more personalized matches Relationship-minded online dating site eHarmony recently upgraded its cloud environment to use CDH and the Intel Xeon processor E5 family to analyze a massive volume and variety of data. The technology is helping eHarmony deliver new matches to millions of people every day, and the new cloud environment accommodates more complex analyses to create more personalized results and improve the chances of relationship success.

Once upon a time, meeting a partner online was not seen as conducive to a happily ever after. In fact, it was seen as a forbidden forest.

However, in the modern age of time poor, stressed-out professionals, meeting someone online is not only seen as essential, it can also be considered to be the more scientific way to go about the happy ending.

For years, eHarmony has been using human psychology and relationship research to recommend mates for singles looking for a meaningful relationship. Now, the data-driven technology company is expanding upon its data analytics and computer science roots as it embraces modern big data, machine learning and cloud computing technologies to offer millions of users even better matches.

eHarmony's head of technology, Prateek Jain, who is driving the use of big data and AI modelling as a way to improve its attraction models, told CMO the matchmaking service now goes beyond the traditional compatibility into what it calls 'affinity', a process of generating behavioural data using machine learning (ML) models to ultimately offer more personalised recommendations to its users. The company now runs 20 affinity models in its efforts to improve matches, capturing data on things like photo features, user preferences, site usage and profile content.

The company is also using ML in its distribution, to solve a flow problem through a CS2 distribution algorithm to increase match satisfaction across the user base. This produces offerings like real-time recommendations, batch recommendations, and something it calls ‘serendipitous’ recommendations, as well as capturing data to figure out the best time to serve recommendations to users when they will be most receptive.

Under Jain’s leadership, eHarmony has also redesigned its recommendations infrastructure and moving over to the cloud to allow for machine learning algorithms at scale.

“The first thing is compatibility matching, to ensure whomever we are matching together are compatible. However, I can find you the most compatible person on the planet, but if you’re not attracted to that person you are not going to reach out to them and communicate,” Jain said.

“That is a failure in our eyes. That’s where we bring in machine learning to learn about your usage patterns on our site. We learn about your preferences, what kind of people you’re reaching out to, what images you’re looking at, how frequently you are logging in to the site, the kinds of photos on your profile, in order to look for data to see what kind of matches we should be giving you, for far better affinity."

As an example, Jain said his team looks at days since a last login to find out how engaged a user is in the process of finding someone, how many profiles they have checked out, and if they regularly message someone first, or wait to be messaged.

Read more: 9 machine learning myths

"We learn a lot from that. Are you logging in three times a day and constantly checking, and are therefore a user with high intent? If so, we want to match you with someone who has a similar high intent," he explained. 

“Each profile you check out tells us something about you. Are you liking a similar kind of person? Are you checking out profiles that are rich in content, so I know you are a detail-oriented person? If so, then we need to give you more profiles like that.

“We look at all these signals, because if I present a wrong person in your five to 10 recommended matches, not only am I doing everyone a disservice, all of those matches are competing with each other."

Jain said because eHarmony has been operating for 17 years, the company has a wealth of knowledge it can now draw on from legacy systems, and some 20 billion matches that can be analysed, in order to create a better user experience. Moving to ML was a natural progression for a company that was already data analytics heavy.

Read more: 4 data analytics trends that will dominate 2018

“We analyse all our matches. If they were successful, what made them successful? We then retrain those models and assimilate this into our ML models and run them daily,” he continued.

With the skillsets to implement ML in a small way, the eHarmony team initially started small. As it started seeing the benefits, the business invested more in it.

“We found the key is to define what you are trying to achieve first and then build the technology around it," Jain said. "There has to be direct business value. That’s what a lot of businesses are getting wrong now.”

Machine learning now assists in the entire eHarmony process, even down to helping users build better profiles. Images, in particular, are being analysed through Cloud Vision API for various purposes.

Read more: Big data analytics: The cloud-fueled shift now under way

“We know what kinds of photos do and don’t work on a profile. Therefore, using machine learning, we can advise the user against using specific photos in their profiles, like if you’ve got sunglasses on or if you have multiple people in it. It helps us to assist users in building better profiles,” Jain said.

“We consider the number of communications sent on the system as key to judging our success. Whether communications happen is directly correlated to the quality of the profiles, and one the biggest ways to enhance profiles are the numbers of photos within these profiles. We’ve gone from a range of two photos per profile on average, to about 4.5 to five photos per profile on average, which is a huge leap forward.

“Of course, this is an endless journey. We have volumes of data, but the business is constrained by how quickly we can process this data and put it to use. As we embrace cloud computing technology where we can massively scale out and process this data, it will enable us to build more data-driven features that can improve the end user experience."

Follow CMO on Twitter: @CMOAustralia, take part in the CMO conversation on LinkedIn: CMO ANZ, join us on Facebook:, or check us out on 






Join the newsletter!


Sign up to gain exclusive access to email subscriptions, event invitations, competitions, giveaways, and much more.

Membership is free, and your security and privacy remain protected. View our privacy policy before signing up.

Error: Please check your email address.
Show Comments

Latest Videos

More Videos

Great piece Katja. It will be fascinating to see how the shift in people's perception of value will affect design, products and services ...

Paul Scott

How to design for a speculative future - Customer Design - CMO Australia

Read more

Google collects as much data as it can about you. It would be foolish to believe Google cares about your privacy. I did cut off Google fr...

Phil Davis

ACCC launches fresh legal challenge against Google's consumer data practices for advertising

Read more

“This new logo has been noticed and it replaces a logo no one really knew existed so I’d say it’s abided by the ‘rule’ of brand equity - ...


Brand Australia misses the mark

Read more

IMHO a logo that needs to be explained really doesn't achieve it's purpose.I admit coming to the debate a little late, but has anyone els...


Brand Australia misses the mark

Read more

Hi everyone! Hope you are doing well. I just came across your website and I have to say that your work is really appreciative. Your conte...

Rochie Grey

Will 3D printing be good for retail?

Read more

Blog Posts

How to design for a speculative future

For a while now, I have been following a fabulous design strategy and research colleague, Tatiana Toutikian, a speculative designer. This is someone specialising in calling out near future phenomena, what the various aspects of our future will be, and how the design we create will support it.

Katja Forbes

Managing director of Designit, Australia and New Zealand

The obvious reason Covidsafe failed to get majority takeup

Online identity is a hot topic as more consumers are waking up to how their data is being used. So what does the marketing industry need to do to avoid a complete loss of public trust, in instances such as the COVID-19 tracing app?

Dan Richardson

Head of data, Verizon Media

Brand or product placement?

CMOs are looking to ensure investment decisions in marketing initiatives are good value for money. Yet they are frustrated in understanding the value of product placements within this mix for a very simple reason: Product placements are broadly defined and as a result, mean very different things to different people.

Michael Neale and Dr David Corkindale

University of Adelaide Business School and University of South Australia

Sign in