Forecasting from Similar Users
Overview
One of the primary methods of forecasting a customers rating of a product is to
find other customers similar to the given customer, and taking an average of their ratings for
the product in question. (see
nearest neighbor method)
Determining Similarity
A central task in the similar user method is to find a function
(a
kernel)
which can compute a measure of similarity between two customers. This function can be based on
- Feature Based
- if a set of features for each customer is available (age, gender, income, education etc....),
then a kernel can be constructed from those features. Of course, determining the form of the function
(finding the optimal function) is the challenge
- Ratings Based
- if the analyst has a set of ratings from the two customers, it is common to calculate a
correlation
among the two sets of ratings using the correlation as the similarity measure. Typically,
a customers ratings are first
normalized
before the correlation is run. In addition, the correlations can only be run on products for which
both customers have supplied a rating.
- Latent Factors