AI Thinks Men Are Shallow

The data doesn’t lie. We started noticing this several years ago, first with some of the attraction data that was available. Check out the raw data for men/women attraction as a function of age:

According to the “human” labeled data, most likely labeled by men, the attraction for a woman steadily declines with age. It is also curious to note that it peaks with fertility. A man’s attraction stays steady after 40, which doesn’t really seem fair.

First Impression:

Last year in Chicago we showed the first results with YouTube’s first impression dataset where the beauty bias for the first impression score was worse than the bias between a white male and black female. This means that attractive people are “more” likely to be hired for both men and women, but for women, the bias is ~30% worse.

ABC’s Bachelor/Bachelorette Model:

For those that are not familiar with the wonderful show, the Bachelor/Bachelorette, it is a popular dating/reality TV show in the US. It is actually kind of terrible, a single man (a.k.a The Bachelor), will have 30 single women date him for the show and then pick a single person to marry. There is crying, jealousy, lying, drama, and everything you would expect with this type of highly competitive/twisted scenario. There is also a Bachelorette version where a woman (the Bachelorette) is chased by 30 testosterone loaded challengers.

Which show do you think was the most predictive on just looks? The Bachelor or the Bachelorette?

You guessed it, the Bachelor is much more predictive using just a face. So that means when it comes to men vs women there are 3 strikes against men.

The 3 Strikes Against Men:

1. Beauty Before Age: Men are more likely to depress perceived beauty with age as seen on the beauty data
2. Hired For Looks: Men allow beauty to influence their hiring more as seen with the first impression dataset
3. Men Are Predictable, Women Less So: Men allow looks to influence their selection (Bachelor vs Bachelorette) more than women. Women are more likely to allow the behavior to impact this. *I think it is important to argue that maybe we are just predicting the difference in show predictability/scriptability from the directors. Do you think the show is scripted or a true reality show? Leave in the wind?

If you are a nerd keep reading. Our bachelor/bachelorette model isn’t JUST a deep-net regressor, it is a holistic deep net regressor from hell. I’ll be upfront and say most data scientists would be intimidated by this thing. So, going into the network isn’t just an image, but it is images, specific face crop zones to drive unique feature development. We then include 13,517 structured variables from previously trained models around gender, age, beauty, country, genetics, ethnicity, etc…

This shows the number of image crops going into our model. Similar to a focal point model in a video game, we are focusing on certain aspects of the face and allow the model to get traction where it needs to. Having this many crops also allows us to dodge some of the issues with max pooling being too aggressive with downsampling features.

I’m tired, but every day we get 10 new comments on this article we will post new/analysis insight. So please like/comment/share and we will keep expanding this technical analysis section with more & more insight/details on our approach (e.g. feature importance breakdowns, face heatmaps, detailed network architectures, raw validation results in pickle files to play with, code snippets, etc..).