Marriages made in heaven with a little help from Bayes

Sriram Sharma May 14, 2018

Imagine you want to find a partner. Depending on the relationship you are looking for, you turn to a Tinder or a Grindr or a You fire up the app. Your phone takes your picture and the algorithms powering the app compare it with millions of profiles of other relationship-seekers online. Using a blend of artificial intelligence (AI) and face recognition techniques, the app shows profiles that you have a higher chance of matching up with.

As you continue using the app, it tracks everything: the pauses you make on certain profiles, the speed of the swipes between them, the time spent on a profile, how your eye moves, the details you read up, whether you hesitate to touch the contact button… . And, with time, say, you have been using it for two weeks, you find that the app gets darn good at making suggestions and you like almost all the profiles it throws up. The app even offers to write up an introductory note to the dog lover you kind of like based on context from his or her social media profile: “Hi. Grieve with you at RCB’s pasting. Tell you wot: will be at a doggie hangout Sat/4PM/CubbonPark. Want to come with Patch? Will be fun.”

Sounds futuristic and dodgy on privacy, even if permissions are taken? Hardly. did exactly that – except the social media scraping and contextual conversations – about two years ago with dating app Tinder. According to its creator Justin Long, Bernie was good at its job. “Bernie brings a more personified, human experience to online dating by summarizing in conversational language and we leave the graphs and charts and stats to the background. Bernie still opens conversations with fun and random questions,” he wrote on his blog. Long claimed a 99.86% feedback accuracy with Bernie. Tinder, without explanation, pulled Bernie off its platform in June 2017.

AI has made tentative moves in the business of matchmaking in the past and as machine thinking gains traction, we are seeing more of that mainstreaming around us. This fortnight, Facebook CEO Mark Zuckerberg announced the social media giant’s plans for a dating service. When its gets unveiled later this year, you can be sure there will be AI and machine learning algorithms at work in the latest Facebook creation. Almost every matchmaking site already claims to use intelligent matchmaking algorithms that pair and recommend user profiles based on a variety of parameters: ones explicitly filled in by the user and ones inferred through the use of machine learning.

Can the inferences of the matchmaking algorithms be trusted?

Algorithmic secret sauce

Machine learning algorithms make all manner of decisions that influence our daily lives: from the best route to take to work to what news shows up on your social feed, the movies you should watch next or the music you’ll enjoy, what you need to buy to determining your job-worthiness. It should be little surprise then, that they also determine who you should be paired with. Deep learning algorithmic techniques can be used to gauge the attractiveness of a person. Earlier this year, researchers from the South China University of Technology, in Guangzhou released a benchmark dataset for facial beauty prediction.

If that seems dystopian to you, take the case of Tinder’s algorithms today (not Bernie), which is known to pair people based on their attractiveness. Tinder says that it relies on automated decision-making or profiling to recommend compatible people based on a number of criteria, including user-provided attributes, common interests, search criteria and location. However, there’s more to their proprietary algorithm than they talk about: the app is known to maintain a secretive internal rating system akin to how the Elo score ranks chess players relatively to one another.

Swipehelper, an explainer for online dating sites warts and all, says that Tinder’s algorithm is used to pair people at the same level of attractiveness, calculated by the number of people who swipe right on you (indicating interest), and their Elo score. This score goes down if you are seen as someone who plays “outside of your league” or if you are seen as too eager – someone who swipes right on all the profiles. Swiping right on over 2,000 profiles in less than hour, using say, a fidget spinner, will also result in a temporary lockout of your account.

Data-driven techniques are used at India’s matrimonial giants – serving the world’s largest single-country young population, defined as less than 25 years old – across a spectrum of customer touch points. Since 2015, for instance, the group has put into production what they call MIMA (’s Intelligent Matchmaking Algorithm), which provides real-time suggestions on profiles appropriate to the user based on mathematical rules and machine learning techniques.

The group with over 300 matrimony websites has a team of around 10 data scientists, says Meenakshi Variankaval, who heads the data sciences lab at the group. The team works on refining its data-driven approaches across the user journey – from customer acquisition, engagement and experience, retention, and value maximisation, she says.

MIMA (’s Intelligent Matchmaking Algorithm) has resulted in a “significant double-digit increase” in users finding matches that end up in marriage, says Meenakshi Variankaval, head, data sciences lab at

“After MIMA launch, we saw a significant increase in our C2C (consumer-to-consumer) metrics and these metrics have seen a steady increase with continuous refinements to the algorithm,” Variankaval says. The algorithm has resulted in a doubling of the chances of a user liking the MIMA-sorted profile and a “significant double-digit increase” in users finding matches that end up in marriage. “The MIMA algorithm has resulted in increased customer engagement and higher user satisfaction, and increased sales,” she says. More details on MIMA were not furnished.

Also See: The life lessons that helped Murugavel Janakiraman on his journey to market domination 

Gourav Rakshit, CEO,, also a top player in matrimonial services, says that its machine learning efforts have had a big impact on core matchmaking aspect:  show users the best matches based on their explicit and implicit actions. “You can use all kinds of inputs on how people engage with the profile, not just who they interact with, but how long they spend on a profile etc.,” Rakshit says. “It starts to go where a human can’t – in terms of setting up heuristic models. They might look exactly the same based on the parametric information they provide us, but have a very different ethos when looking for a life partner.”

Machine learning techniques are used to add inferential inputs on a profile, when users leave out some bits of information.These inferences could be around the income level of the individual, whether the individual is well travelled or not. Or, it could be things the platform doesn’t explicitly ask as parameters but are necessary to be able to find compatibility in matchmaking. “It allows us to do missing field analysis, where we’re able to use the rest of what we know about a person to try and create a picture of the person– to be able to create the right matches or the right people that this person should be shown to,” Rakshit adds.

Data scientists tasked with growing some aspect of their business by 10-15% every quarter, says Gourav Rakshit, CEO,

He wasn’t forthcoming on details citing intellectual property concerns. He limited himself a conceptual overview: “Basically if you can identify a cohort of people who exhibit that behaviour, then you can expand it to a larger cohort that is similar to these people. That’s broadly the technique for all the inferential stuff that we use.”

A team of six data scientists at is tasked with growing some aspect of the company’s business by 10-15% every quarter. “In the last few quarters, I’m pleased to say that they’ve demonstrated that value, and it will be an area that we will continue to invest in,” says the CEO.

Testing anti-fraud algorithms is also known to use Amazon Rekognition, a deep learning-based image and video analysis service that can screen photos to check for nudity, blurred photos, celebrity photographs, and other image-related hygiene issues. says that it has cut down manual work related to photos by 50% and the time it takes for a photo to appear on the site by 95%. also says that it uses machine learning to ensure that the members registered are genuine and serious about finding a life partner. “At the time of registration, we have an alert-driven validation process where photo and personal information are scanned for inconsistencies and false information,” says Variankaval. “Our algorithms have identified personas of ‘Fraud’ based on history and we use it to identify profiles that look ‘suspicious’ for validation and close monitoring.” However, they declined to answer our follow-up questions on whether they use computer vision algorithms to do this, or mention how many accounts were flagged as fraudulent since its deployment into production. The team declined to answer what parameters are used and whether facial recognition algorithms are employed to match people or identify fraudulent profiles.

FactorDaily tested both and on a simple computer vision related problem: by creating two dummy profiles on each site using the same profile photo. Did they flag the profiles as duplicate? Both failed and all the four accounts were still active six days after creating them. The companies did not reply to a request for comment specifically on this issue.’s spokesperson asked us to refer to their previous statement that appeared in a FactorDaily feature.

“We suspend the (fake) profile. We encourage the aggrieved person to file a police complaint and we support the investigation by providing details. When we are requested by a competent authority, we provide the alleged defrauder’s details to them,”  Murugavel Janakiraman, founder and CEO of

AI and machine learning, it appears, can trip up at times. It might seem like we’re nitpicking but matrimonial fraud is a serious problem as we described in detail in the case of Tanmay Goswami, a serial fraudster who used matrimonial sites to con and defraud several women – eight of who have filed police cases alleging embezzlement of a cumulative Rs 1.25 crore. The good old ‘Trust Badge’ that uses the user’s social media profiles and even a government ID for verification may work best.

A new other half

Launched in March this year,, which calls itself “India’s first AI-powered life partner search experience”, takes a slightly different approach to matchmaking, by putting a personality test before every user at the time of registration. It requires a LinkedIn account to register, and users are encouraged to verify their profiles with their work email, and government ID.

Pawan Gupta (left) and Rahul Namdev (right), co-founders of, which calls itself an “AI based partner prediction engine”.

Through a set of 16 Likert scale questions, the startup maps a user’s personality across six dimensions: emotional, social, intellectual, physical, relationship, and values. (The Likert scale measures people’s attitudes to a topic.) It asks the user to evaluate oneself on a scale of 1-7 on positive and negative questions like “How kind are you”, “How lively are you”, “How dominant are you”, “How quarrelsome are you”… .

This initial score of a user is continually updated based on behavioural information such as click-and-scroll activities and time spent on different sections of the profile of a suggested match. The personality prediction scores are further updated based on private feedback about a user from his or her matches after interactions. These scores will be further refined through gamified personality questions.

A screenshot of the user feedback feature on, used to update their personality predictions.“We use recursive probabilistic Bayes update algorithm to estimate and learn about the correct personality of users over time,” says Betterhalf’s co-founder and CTO Rahul Namdev. Bayes refers to Thomas Bayes, an 18th century statistician best known for a theorem named after him and his probabilistic models.

“Based on analysis of thousands of data points captured through our 16 likert type questions, we first assign a initial probability (called prior probability) to the reliability of each of these answers. Based on these probabilities, we estimate an initial probabilistic personality score,” he explains. “As we keep on learning more about the user through gamification and through private feedback from others, we keep on updating this probability score (posterior probability) and personality score.”

Betterhalf uses bayesian machine learning methods to estimate compatibility between individuals.

Compatible matches are filtered and ranked, based on partner preferences of the users through partitioning-based clustering methods. “We define a compatibility distance matrix – which is the weighted average of euclidean distance within each cluster to find out all the matches which are most compatible with you,” says Namdev.

Happily Ever After?

“Currently at a user base of over 2,000, we’re around six months from our product-market fit,” says Betterhalf co-founder and CEO Pawan Gupta. The platform claims to have made over 200 matches so far. Gupta and Namdev are both alumni of the Massachusetts Institute of Technology.

“The best thing about Betterhalf is that they get us logged in from LinkedIn. That helps with verifying the profession of wherever they are working. Most people don’t lie on LinkedIn,” says a Pune-based software professional, who requested anonymity recounting her story of finding a life partner online.

In the three-four years her parents and she searched for a partner, she encountered several profiles that were fake. One experience on in December 2016 was particularly traumatic. “We found a match and the guy spoke to my dad and he was like, he works for Google, and he’s in HR over there, and he’s a friend of Virat Kohli, and all that…,” she says. “We (later) got to know that the photographs he shared was of someone different from his. He called himself Chirag, the photo used was of a guy called Geet. There was no HR in Google India with either of their names. It took six months to find out the real story.”

A screenshot of the user feedback feature on, used to update their personality predictions.

Still, the fraudster conned her out of around Rs 30,000 in all. “He asked me to gift him a One Plus 3, and I was stupid enough to do so,” she says. She provided us screenshots of bank transfers and the address the phone was mailed to.

In a response to a request to for comment, they said that they use AI-backed algorithms to flag suspicious users, and a team of executives to go through flagged cases.

She is now engaged to a person she met on Betterhalf.

Manoj Singh, another Pune software professional employed with Infosys, found his partner on Betterhalf. He had signed up for a free beta after hearing about the site on an online bulletin board at work.

“Someone from Infosys had posted about this site, when a beta version was launched,” he says. On the site, Singh was curious about “intelligent matchmaking” and used the facility to book a video call with the CEO. “Pawan mailed details about his algorithm and I came to know that he’s focusing on compatibility as well, which the family doesn’t bother about. There were a lot of questions about how I think. It was evaluating me on my emotional and behavioural values,” recalls Singh.

On Betterhalf, Singh found the profile of a woman he had met formally with his family earlier. “It was just a half an hour meet… I felt she’s not my kind of a match,” he says. “I checked my compatibility with her. It was around 80%, which was very high. I sent a request to her and got a chance to talk to her separately. After talking to her again, I realised that she’s a good match for me.”

He got married in April. Did the Betterhalf algorithm do a good job of measuring compatibility between potential partners, we asked him. “Till now, things are fine, maybe you should call me after two years,” he laughs.



Disclosure: FactorDaily is owned by SourceCode Media, which counts Accel Partners, Blume Ventures and Vijay Shekhar Sharma among its investors. Accel Partners is an early investor in Flipkart. Vijay Shekhar Sharma is the founder of Paytm. None of FactorDaily’s investors have any influence on its reporting about India’s technology and startup ecosystem.