User profiling with Mahout from categorized user behavior

Question

I'm trying to cluster and classify users with Mahout. At the moment I am at the planning phase, my mind is completely mixed with ideas, and since I'm relatively new to the area I'm stuck at the data formatting.

Let's say we have two data table (big enough). In the first table there are users and their actions. Every user has at least one action and they can have too many actions, too. About 10000 different user_actions and millions of records are in the table.

user        - user_action
u1          - a
u2          - b
u3          - a
u1          - c
u2          - c
u2          - c
u1          - b
u4          - f
u4          - e
u1          - e
u1          - d
u5          - d

In the other table, there're action categories. Every action may have none or multiple categories. There are 60 categories.

user_action - category
a           - cat1
b           - cat2
c           - cat1
d           - NULL
e           - cat1, cat3
f           - cat4

I'm going to try to build a user classification model with Mahout but I've no idea what I should do. What type of user vectors should I create? Or do I really need user vectors?

I think I need to create something like;

u1 (a, c, b, e, d)
u2 (b, c, c)
u3 (a)
u4 (f, e)
u5 ()

Problem in here, some users performed more than 100000 actions (some of them are same actions)

So; this is more useful, I think;

u1 (cat1, cat1, cat2, cat1, cat3)
u2 (cat2, cat1, cat1)
u3 (cat1)
u4 (cat4, cat1, cat3)
u5 ()

The things I also worry about are

• How should I weight categories for users? For example u1 has at least three action that related with cat1, while u3 has only 1. These one should be different?
• How can I decrease the difference between active users and passive ones? Like u1 has too many actions and so categories, u3 has only 1.

Any guidance are welcome.

Answer 1

I think one solution for this might be using Matrix factorization, to perform some kind of collaborative Filtering. This way you will have not have to deal with the two concern points that you have, explicitly.

To do this, create a [User X Action] Matrix W, where rows are users and columns are actions. W(i,j) i.e. the Matrix entry on i row and j column would be the number of times the user i has performed action j. The matrix entries where no data is given are treated as missing data values by most collaborative filtering algorithms.

When you factorize this Matrix, you factorize it into two parts W= A*B. The A matrix will have rows are users and columns as latent dimensions. From this A, you get a feature vector for each User. You can then run clustering on these feature vectors to cluster the users. For more intuition about the factorization approach, look up "Matrix Factorization for collaborative Filtering" on google. Mahout has matrix factorization implementations that you can use.

Using the action categories is another matter altogether. It depends on what a category means. Does it mean that every action inside a category is very similar to each other or that they just have one or two aspect of similarity. A very easy approach would be to factorize a [User X (Action + Category)] Matrix. This can be solved using the same implementation mentioned earlier. You can also have more fancy schemes, like fixing part of the Matrix B to be very similar to the [Action X Category] matrix. But these schemes will require you to write optimization code yourself.