Difference between Hybrid CF Model and Learning-to-Rank Model

Hi,

I’m a little confused when looking at the examples in the documentation. What is the difference between the Hybrid and Learning-to-Rank models implemented here?

From the documentation, a Hybrid model is defined by:

model = LightFM(loss='warp',
                item_alpha=ITEM_ALPHA,
                no_components=NUM_COMPONENTS)

A Learning-to-Rank model is defined by:

warp_model = LightFM(no_components=num_components,
                    loss='warp',
                    learning_schedule='adagrad',
                    max_sampled=100,
                    user_alpha=alpha,
                    item_alpha=alpha)

The default lightFM model is:

LightFM(no_components=10, k=5, n=10, learning_schedule=’adagrad’, 
loss=’logistic’, learning_rate=0.05, rho=0.95, epsilon=1e-06, item_alpha=0.0, 
user_alpha=0.0, max_sampled=10, random_state=None)

I’m having trouble seeing a difference between both models. Does LightFM differentiate between the Hybrid CF and Learning-to-Rank? How do I make sure the model that I’m creating is a Hybrid model or a Learning-to-Rank model?

Would appreciate any advice on this. Thank you!

[maciejkula]

1. If use use the warp or bpr losses, your model is a learning-to-rank model.
2. If you have any features in your item or user feature matrices (other than the identity features per item/user), your model is a hybrid model.

[SimonCW]

I’m closing this issue because it has been inactive for a long time. If you still encounter the problem, please open a new issue.

Thank you!