Clarification on input and output formats

[dglodzik]

Hello there,

I am really excited to test the software. I work on mutational processes in cancer and the model has some of the properties I need.

I wonder if you could help by clarifying the input and output formats. I managed to install and run the example, but don't fully understand the formats. Would you help by pointing me to a description of the input and all the files that the program outputs? I am mostly interested in exploring the components, a bit like you did with the movie preferences in the article, and trying to understand common themes.

Kind regards
Dominik

[jhofman]

the basic format for observations are (user_id, item_id, rating) tuples in the train/validation/test files. auxiliary files provide summaries of these observations.

for instance, items.tsv is a count of how many times each item is rated, and similar for users.tsv.

hope that helps in getting started.

[dglodzik]

Thank you!

Could you point me to which file in the output has beta's_{i,k} ? I imagine that would be a matrix of the size I,K, where I is the number of items and K is the number of components.

Would it be an issue if the rankings for different movies wee on drastically different scales? For example, it one movie was ranked 0-10000, whereas another 1-30?

Dominik

[slinnarsson]

After fitting the model, beta is in hbeta.tsv and theta is in htheta.tsv. In these files, the first column seems to be a serial number, the second is the movie/user ID, and subsequent numbers are the k components.

Here's how to parse the files into Python matrices of shape (n_users, k) and (n_movies, k):

theta = np.loadtxt("htheta.tsv")[:, 2:]
temp = np.loadtxt("hbeta.tsv")
beta = temp[:, 2:][np.argsort(temp[:, 1]), :]  # the beta matrix with the correct rows ordering

Note that you need to sort beta, since the rows are not in order. Theta seems to always be in order.