Developing KNN with custom distance matrices

Question

I am trying to predict the extent of suitability (0.5 to 3.5) for a section of non-fiction books based on a few parameters. This is for a set of subscribers that we want to predict books suitability: based on score I plan to display star ratings: most suitable(2.5 to 3.5), mildly suitable(1.5 to 2.5), not relevant (0.5 to 1.5).

Example dimensions:

• price: continuous

• subject: categorical (humor, adventure, mystery, sports, non-fiction)

• hardbound: binary

• recency of book: categorical

We initially start with a default profile for all user and after getting feedback from him (as 1, 2 or 3 stars), build a customized profile.

Given that we will initially start with small volume of user data, looking to do a k-NN to identify nearest neighbors on the dimensions, and calculate suitability score based on weighted suitability of those neighbors.

Given the categorical data, I wanted to create custom distance matrices so that I could provide the relative distance between categories. (e.g. to indicate that humor and adventure are close preferences; sports and nonfiction are far apart)

Two questions:

1. Is my current approach valid- any feedback or pointers?
2. What specific k-NN implementation (package) would allow me to provide my custom distance matrices?

Answer 1

I would personally take a look at collaborative filtering as this will take into account the information of the user itself, but also the information of similar users. If let's say, my best friend and I have the same taste in books, knowing his taste can help you guess mine! I'd recommend watching the Andrew Ng lectures on Coursera about recommender systems.

KNN might seem like a natural choice if you don't know about recommender systems (I have been there!) but please, make sure to read about them before you decide to use KNN.