If your dating app said:
“Our AI says that you two will stay together as a couple with a 79.3% certainty.”
Would you wonder how they came up with that number?
Valentine’s Day and Artificial Intelligence
As Valentine’s day approaches, the topic of dating and with that, dating apps, creeps up in people’s minds. Some have used it, some have no interest, and some might be curious about using it. Irrespective of people’s opinions on dating apps, I find that an interesting topic here is that of the algorithms behind the scenes. The math, or lack of sometimes, behind the recommendations people see when interacting with these apps. As a data scientist, there are many things one has to look at when working with a dating app. In the past, I have had experience in social networking apps where the purpose was to recommend people that should connect with each other.
AI Matchmaking Scenario #1
The first and simplest way to approach the problem is to treat it like a simple optimization game. A data scientist can look and think about the product as a proposal from the software and a response from the user. The proposal being: “Would you like to connect with person X?” and the response is a yes or no from the user. When approached this way, things are very simple for a data scientist. They are just trying to suggest people that will elicit a yes response; this means they made a good suggestion.
Let’s look at that more closely. A data scientist can collect data of all previous suggestions the app has made and which people clicked yes or no to which suggestions. Then a data scientist will collect information about the people being shown and the ones clicking yes and no, and that information is what data scientists call features. This data set will now allow a model to be trained to predict the probability of someone clicking yes or no when a match is suggested. A data scientist would be happy to see that the model achieved high accuracy, however, that is not the end of the story. What the model probably learned is that super attractive people will usually receive a yes click. This is not necessarily what is best for a social or dating app.
AI Matchmaking Scenario #2
Let’s imagine a slightly different situation. Let’s imagine an app that recommends movies for you to watch. If that app is very good and most times it recommends movies people actually do follow the recommendation, that would be a successful app. Here is the big caveat in the differences between the two examples. With movies, the movie does not have to like you back. If you watch the movie and like it, that’s it; not so with people. Also, no movie will be sad if no one watches it, and lastly, no movie will get annoyed if it has 1 million requests to be watched. You can see how those three things would be problems for people in the dating app scenario? When working with a dating app, and building a machine learning model for it, you really need to understand the business. The understanding of what success means for your customers is infinitely more valuable in this case than knowing how to tune a neural network or the latest and greatest machine learning algorithm.
This example goes to show that equipping the right person with the right tools can lead to a big impact, creating a better business. In this case, the best person for the job is the domain expert, and the tools needed are data science and machine learning tools. This is all part of what I believe is the evolution of data science and the roles within. As the technology continues to advance and become more accessible, the involvement non-data scientists will have will grow, and the role of today’s data scientists will shift. I will be writing other articles on the topic soon.