Integration of User- & Item-Features

Hi,

I have a problem with integrating user- and item-features. After some research I now know that I have to concatenate my user/item-feature matrix with an identity matrix. After doing this, my user-matrix looks the following:

When I know integrate this matrix in csr-format into my model fit it with (user_features = my_features) export the top-N recommendation using the predict function (with user_features agian!) I get the following precision when evaluating it. The blue line here is without adding User_features!

What is going wrong there? 😦

My code looks like this:

NUM_COMPONENTS = 10
NUM_EPOCHS = 20
learn_rate = 0.02
NUM_THREADS = 10

model = LightFM(loss = 'bpr, no_components=NUM_COMPONENTS, learning_rate=learn_rate)

model.fit(train, epochs=NUM_EPOCHS, num_threads=10, user_features=user_features)

` after doing this I export the top-N for every User in my testset:

recommendations = list()
n_users, n_items = test.shape
for user_id in np.arrange(n_users):
     scores = model.predict(user_id, np.arange(n_items), num_threads=10, user_features=user_features)
     top_items_list = np.argsort(-scores)[:20].tolist()
     productids = itemgetter(*top_items_list)(products)
     userid = customers[user_id]
     recommendations.append({'UserID': userid, 'Empfehlungen': productids, 'Scores': np.sort(scores)[::-1 [:20].tolist()})

Here customers are the real labels of the customers.

[maciejkula]

This looks fine, as far as I can tell. Try lowering the learning rate!

[KJenia]

hi! I couldn’t understand one thing, after model.fit could I make model.predict for users_id which was’n in model fit? for example, I fit the model for user_id = [1,2,3] and I want make prediction for user_id=4, is it possible do it without retrain the model? p.s. of course there is no new user or item features for predicted user thanks for suggestion