(Part 2 of 2)
It starts with an algorithm
At Say Allo, we believe the goal of a dating application should be to find quality candidates and empower you to make educated decisions on your swipes.
We started with an algorithm and developed our app around it.
We consider ourselves a hybrid to the aforementioned dating app styles:
We believe the best reflection of a user is in their actions. In this case, that is their likes and dislikes. Rather than rest on our laurels with our questionnaire, we continuously adapt based on the user’s input to the system.
We use Machine Learning to help our users find their most compatible match, starting with baseline inputs based on psychology and statistics and continue to learn based on their swipe history. Our algorithm will improve the more you swipe and provide real time updates to your swipe queue as we continuously learn.
So, to feed our process, we needed to define the inputs. After much development, experimentation and testing, we've collected inputs to feed each thing we learn.
Here’s a short list of what we will learn as you use our app:
1. We learn physical attraction.
All of our users’ photos are analyzed for key indicators that help us understand what attracts you to them. Input varies from skin and eye color to ratios of facial points. We’ll take whatever we can find from the submitted photos and use the available data as inputs.
But as the famous saying goes “Beauty is in the eye of the beholder”.
There’s some argument over whether the Golden Ratio works for attraction on faces and, while we consider a baseline for attraction, we by no means define it for you. You do the swiping and we can discover what you find attractive.
2. We learn compatibility.
Being a swipe app, physical attraction will always largely dominate the actions of a user. However, we feel it necessary to ensure that our users can also see how compatible they are with the candidate matches, providing a baseline questionnaire that we ask them to fill out when signing up and then learning the relationships between their compatible attributes with those they will swipe against.
Swiping is generally a fast action, but taking it slower and reading up about your potential match can help us acknowledge your efforts and tailor your experience based on those inputs. We don't ask hundreds of questions. We ask a dozen of you and learn the relationships ourselves.
3. We learn what’s most important to you.
Swiping patterns are different for all users: While some people will swipe on any pretty face, others will take more time to process what they have in common with the user. Various patterns of swiping must be recognized for any system such as ours to work.
But this is where we might teach you a thing or two about yourself. Perhaps you are physically attracted to a specific face and body, but will still swipe left if that person, say, doesn't like cats and dogs. Perhaps you prefer to date people who share your same political views or religion but that special look of somebody will make you ignore all that. Or perhaps physical attraction is not important to you at all and you just want to find somebody who loves all the same things that you do.
We don't tell you that you should find either attraction or compatible traits more important.
We let you tell us.
There are some companies using collaborative filtering to fulfill their promise of a more intelligent dating application. It's not a terrible approach – one where you are shown certain profiles because other people with similar swipe patterns also liked them (think the Amazon.com "Customers also bought this..") – but that is not what we at Say Allo are doing. We don't want to learn about anybody who might be like you. We want to learn about you.
So, please give us a shot, help us build up our user base and we'll help you find that special someone so that you'll make us the last dating app you'll ever need!
Technical Co-Founder of Unpack'd Technologies
(Part 1 of 2)
Dating apps. If you’ve been single for any period of time in the last 5 years, there is a good chance that you’ve tried one. The model makes sense. They provide ease and security in the dating world that didn’t previously exist. And there are a lot of options, with offerings from industry giants to caterers to a small niche, each providing their own hook to appeal to users like yourself.
I’ve been married for over 15 years so I never needed to try any app for the goal of finding a partner — hence the exploration into each app was wholly a new experience to me.
Despite some variance in feature sets, the majority of the apps follow rather similar approaches to finding a match:
Both approaches are entertaining and can undoubtedly work – most of us have friends who have found love through online dating apps — but after playing with a couple dozen, I question whether these apps are effective in achieving the goal in finding the perfect partner and if that even is the goal. I’ve concluded the answer to be very few for either question.
So it is easy to understand why these apps have started a craze. It is difficult to understand what they’ve done since in light of the accessibility of modern advancements to help achieve the previously stated goal.
Trying to learn… something
Similar to online dating, the surge of Machine Learning in applications and software in the last 5 years is impossible to miss if you even have a passing interest in technology. Of course, learning models have been around for a long-long time. As a veteran of the game industry, there has been a lot of research and application of techniques used in Machine Learning and AI and have been for years. But it is the affordability and improved performance that has made it widespread available — Google, Amazon and Microsoft all provide powerful frameworks and algorithms to solve all sorts of problems.
Dating apps inherently collect a lot of information about you: your stated preferences at the beginning of the app and your learned preferences as you start interacting with the app. However, when running through these dating apps, it is difficult to ascertain what level of advanced learning is being done while using one of them. These apps, at least the big ones, certainly must implement learning algorithms or are at least trying to, as is evident by the numerous job postings for Data Scientists and Engineers with Machine Learning backgrounds as well as the occasional blog post or small article. Yet still, from the outside looking in, it is challenging to evaluate how they are applying learning and whether it is for the purpose of helping the user date smarter.
For the question-heavy sites, such as Match, OKCupid or E-Harmony, you spend a lot of time to onboard, filling out multiple choice scruples, and that data is assumed to help find the best match. Are matching similarities a true indication that you’ve found somebody compatible? Is this machine learning or statistical matching based on pre-conceived notions? When we’re talking about a swipe dating app, the model is normally all about volume: swipe a lot and then hopefully match a lot based on some basic preferences. The user is engaged in performing their own brute force algorithm, looking at many profiles that are clearly not going to be a match or suspicious attractive profiles that clearly won't match with you. Some apps will ask the odd question, but these apps are generally about making you, the user, continue the cycle of swipe/match/date without much insight into whether the effort is meaningful. And if you are using some of these apps to find your soul mate, the swipe mechanism is fun and the process could eventually work. Seeing that it is a strong model for entertainment, where’s the incentive for them to do better?
Of course, there is a contingency who prefer this approach. Swiping is, after all, fun and no variation of this model should discount that. But what about the people who are looking for genuine lasting connections and don't want to sift through endless profiles that are unappealing or clearly tailored to keep you swiping? The culture of online dating has changed dramatically with the swipe app. And with a younger demographic and a different goal in mind, many prefer casting a wide net and seeing what they can catch. And there’s nothing wrong with that — but knowing that, why would apps who support this need to invest the time and effort to change their model to support those who want to just find someone perfect?
Unlike other industries, there’s little publicly available about each company’s inner workings when it comes to their process so it is perhaps unfair to assume that not much is being done. But, for fun, take 5 minutes and start googling for Machine Learning and your favorite dating site. There’s not a lot available. A lot of user speculation and not much affirmation from the company’s themselves.
Overcoming the short-comings
There is one other very popular result that comes up when doing these web searches about Machine Learning and dating apps: people interested in finding matches on these sites creating outside-the-box solutions to maximize their potential to do so.
A successful app like OKCupid will ask the user hundreds of questions up front to load their matching engine to try and find the most like-minded profiles. The wired article “How a Math Genius Hacked OKCupid to Find True Love”  relates a common problem with these sites — they generate an immutable score between users based on these questions. If there is a problem with the questions, as there are in the case of the article, then the user’s experience may be limited based on that first assessment.
Tinder, being the leader in the swipe dating app industry, has a large user base and an API that is possible to received access to. And as a swipe-based app, they require the user to swipe through as much as they can. There have been many attempts to create tinder bots that try to do all the heavy lifting, sifting through the numerous profiles on behalf of the user. These generally train a model on what they believe to be attractive and then root out all the attractive users for right swipes, occasionally starting the conversation immediately.
And yet, there are plenty of stories online of men and women trying to force Tinder to work for them based on this API. They want to find their match, but their only option is to run some automation based on a pre-defined ruleset. They need to play Tinder’s game.
In my next post in this series I will dive into greater detail about the learning aspect of our algorithm, and uncover some of our “how” behind our machine learning application! In the meantime, we hope that you download Say Allo and start learning something new about yourself, and somebody new!
Technical Co-Founder of Unpack'd Technologies