Prediction and explanation of user rating based on multiple criteria

Question

I'm trying to figure out a way to both predict how a user would rate a certain document, as well as an explanation of why certain documents are rated a certain way.

A user is represented by:

• favorite topics;
• average length of the documents the user read;
• the distribution of the times of day a user reads the documents;
• location;
• age.

With this information I want to:

1. Predict how much a user will like a document (based on the document's topics, length, the time of day, etc)
2. Get explanations like: documents with this length are usually read around this time; or, documents about these topics are usually read at these locations.

I hope this question is not too broad. I just need some starting tips.

Answer 1

Since (per your comment) you have a rating from 1 to 10, I think the first thing to try is ordinal logistic regression. Another possibility that I have been reading about is beta regression, but this will be less familiar to nearly everyone (although, from what I'm reading, it looks much more flexible).

Answer 2

As Peter Flom suggests, logistic regression would be a good place to start.

If your goal is automated prediction more than an easily interpretable model, you might want to read through the solutions thread for Kaggle's Event Recommendation Engine contest. That was focused on a binary interested/not outcome, but otherwise has some similarities to your problem.