What is the right way (format) to input user features to the build_user_features function?

I have been struggling for a while now trying to input my user features into lightfm to try to include them in my recommendation model. I have read many other issues refering to similar problems, but I still can’t manage to work solve my problem.

I have my data in a pandas dataframe. My user IDs are strings like "AHS-1", and I’m trying to include one user feature to start. For what I’ve read, the format in input user features into the build_user_features function is:

[user_id1, [user_feature1, user_feature2], [user_id2, [user_feature1, user_feature2] …]

I’ve tried many options to create this, the latest being creating a dataframe with the user IDs and the feature, then converting this into a tuple, but I get the error “TypeError: ‘int’ object is not iterable” since my feature is an int.

Here’s my code:

user_features_pd=pd.concat([user_data['mber_id'],user_data.iloc[:,5].astype(int)], axis=1) 
tuples = [tuple(x) for x in user_features_pd.values]
user_features = dataset.build_user_features((tuples))

I would appreciate very much any help!

[Med-ELOMARI]

well , the error explain itself (Feature 62 not in eature mapping. Call fit first) , i think you used just the function that i mentioned . i recommend you to use the whole class , that will fit every unique existing feature before building it .

let’s pic a super 😄 small example , here we have this interactions :

	item_X	item_Y	item_Z
user_A	0	5	1
user_B	1
user_C		5

with this users details :

user_A	user_feat1	user_feat2
user_B	user_feat3	user_feat4	user_feat2
user_C	user_feat1	user_feat4

and items details :

item_X	item_feat1
item_Y	item_feat2	item_feat3	item_feat4
item_Z	item_feat1	item_feat3	item_feat4

so the right way or format to present the data is :

interactions = [
    ("user_A", "item_X", 0),
    ("user_A", "item_Y", 5),
    ("user_A", "item_Z", 1),
    ("user_B", "item_X", 1),
    ("user_C", "item_Y", 5),
]
users_features = (
    ["user_A", ["user_feat1", "user_feat2"]],
    ["user_B", ["user_feat3", "user_feat4", "user_feat2"]],
    ["user_C", ["user_feat1", "user_feat4"]],
)
items_features = (
    ["item_X", ["item_feat1"]],
    ["item_Y", ["item_feat2", "item_feat3", "item_feat4"]],
    ["item_Z", ["item_feat1", "item_feat3", "item_feat4"]]
)

first we create a Dataset instance

from lightfm.data import Dataset
dataset = Dataset()

then we must fit all the data you have , users , items , all features this step is essential because the LightFM model understand numbers not strings , so we will need to map each string we have to a number , and that’s what dataset.fit do

dataset.fit(
    users=["user_A", "user_B", "user_C"],
    items=["item_X", "item_Y", "item_Z"],
    item_features=["item_feat1", "item_feat2", "item_feat3", "item_feat4"],
    user_features=["user_feat1", "user_feat2", "user_feat3", "user_feat4"],
)

you can see the feature_mapping is a dict of mappings for our data

then we can build the rest (assuming everything we will add here is already fitted (Mapped) in the last step) :

(interactions, weights) = dataset.build_interactions(interactions)
user_features_list = dataset.build_user_features(users_features)
item_features_list = dataset.build_item_features(items_features)

then we can feed our model

model = LightFM(no_components=24, loss="warp", k=15)
model.fit(
    interactions=interactions,
    sample_weight=weights,
    item_features=item_features_list,
    user_features=user_features_list,
    verbose=True,
    epochs=10,
    num_threads=20,
)

to predict , we give the mappings as input , not the strings (user_A …)

print(model.predict(1, list(range(3))))

Results [-0.15689075 0.10851561 -0.19980735]

hope i makes it clear now , ask if not 😵 good luck 🍀

[clementechiu]

Ok thank you very much @Med-ELOMARI ! I think my problem is probably that I was inputting my continous variables wrong (not as a dictionary with weights as suggested in an answer of the issue #433). I tried splitting it into chunks as you suggested and it works. Thank you very much for your help!