Selecting the right algorithm for match probability prediction

Question

Looking for assistance kick-starting a new machine learning scenario. In this case I need to pair one entity (ex. person) with a group of entities (ex. other people) given a history of matching patterns over time.

For example: Assume we have people who like other people with a combination of traits such as height, weight, education level, etc... This notation of 'like' is bi-directional when calculating probabilities. When a new person arrives I'd like to isolate a group of people (ranked by probability) to recommend them too.

In this case the input would be a single person with all their traits and the output would be a list of people with ranked probabilities of matching. There seems to be two main sets of data

• Known people
• Traits known people have liked or not with the traits of people they have liked or not

I am relatively new so not sure which types of training model to use. For example, is this something that can be accomplished most reasonable with a linear regression model, binary classification, boost-tree's, matchbox, etc, etc, etc..? Possibly I need a combination.

I am using the Azure Machine Learning studio which has lots of options but I am just not sure where to start in this scenario. Any suggestions or pointers would be appreciated in helping

Answer 1

It seems a variant of Movie Recommendation problem. Take Netflix for example, they have user ratings for different type of movies; in your case users are telling what attributes they like. When a new user joins and rates the things he/she likes, you will find the people with similar interests, again with Netflix analogy it means recommending movies based on people with same tastes as the new user.

Read on Collaborative filtering, which is used for building recommendation systems. The reason I drew the Netflix/Movie Recommendation analogy is because you can find a lot of literature on it.