Improve matching algo with an autoencoder

[duogenesis]

Potential extra features:

• Number of questions answered
• Vectors from an image model like VGG
• Bag of vectors of bios
• Gender
• Preferred gender
• Age
• Whether one's age is private -- Female users seem to be having trouble with older men messaging them with their age hidden
• Whether one's location is private
• Whether their email domain address from a set of known good ones

Consider using a training objective which provides a bigger reward for harder-to-predict questions. The reason to pick an objective like that is to deal with the fact that some personality traits might have more questions which measure them than others. e.g. there might be 10 questions measuring introversion/extroversion but only 1 for career-focus. You should also consider more heavily weighting the questions people set as search filters because these are likely to be important to most people.