meikuam
commented
3 years ago lightfm uses categorial features, so you should convert numerical to categorical and use it like "tag" for item or user. I will give to you simple toy example of data prepare/train pipeline:
from lightfm.data import Dataset
# your user ids
user_ids = [0,1]
# all available features
user_feature_list = [f"age_{age}" for age in range(30, 40)]
user_feature_list += ["male", "female"]
# your mappings user ids on features
user_features = [
(
0,
{
"age_35",
"male"
}
),
(
1,
{
"age_31",
"female"
}
)
]
# your item ids
item_ids = [1,2,3,4]
# your features
item_features_list = [f"feature_{f}" for f in range(10)]
item_features = [
(
1, {
"feature_0", "feature_1"
}
),
(
2, {
"feature_1", "feature_2"
}
),
(
3, {
"feature_0", "feature_3"
}
),
(
4, {
"feature_9", "feature_1"
}
),
]
# create dataset class
dataset = Dataset(item_identity_features=False)
# add mappings
dataset.fit_partial(users=user_ids)
dataset.fit_partial(items=item_ids)
dataset.fit_partial(user_features=user_feature_list)
dataset.fit_partial(item_features=item_features_list)
# build sparce matrices with
dataset_user_features = dataset.build_user_features(user_features)
dataset_item_features = dataset.build_item_features(item_features)
# build interaction matrices (pairs of (user_id, item_id))
x_train = dataset.build_interactions([
(0, 1),
(1, 3),
(1, 4)
])
# so you can get mappings
[
dataset_user_id_mapping,
dataset_user_feature_mapping,
dataset_item_id_mapping,
dataset_item_feature_mapping
] = dataset.mapping()
"""
(I paste here outputs from jupyter notebook)
# here mapping your ids to internal representation
dataset_user_id_mapping = {0: 0, 1: 1}
# here mapping of user features to internal representation (0, and 1 keys here as we leave user_identity_features param in Dataset class by default)
dataset_user_feature_mapping = {0: 0,
1: 1,
'age_30': 2,
'age_31': 3,
'age_32': 4,
'age_33': 5,
'age_34': 6,
'age_35': 7,
'age_36': 8,
'age_37': 9,
'age_38': 10,
'age_39': 11,
'male': 12,
'female': 13}
dataset_item_id_mapping = {1: 0, 2: 1, 3: 2, 4: 3}
dataset_item_feature_mapping = {'feature_0': 0,
'feature_1': 1,
'feature_2': 2,
'feature_3': 3,
'feature_4': 4,
'feature_5': 5,
'feature_6': 6,
'feature_7': 7,
'feature_8': 8,
'feature_9': 9}
dataset_user_features (sparce matrix with 2 rows - users and cols - user_features, user_features are correspond to dataset_user_feature_mapping)
<2x14 sparse matrix of type '<class 'numpy.float32'>'
with 6 stored elements in Compressed Sparse Row format>
dataset_item_features (same as user_features)
<4x14 sparse matrix of type '<class 'numpy.float32'>'
with 12 stored elements in Compressed Sparse Row format>
x_train (2 sparse matrixes first - interactions, second - weights)
(<2x4 sparse matrix of type '<class 'numpy.int32'>'
with 3 stored elements in COOrdinate format>,
<2x4 sparse matrix of type '<class 'numpy.float32'>'
with 3 stored elements in COOrdinate format>)
"""
# then we create model and train on toy data
from lightfm import LightFM
model = LightFM(no_components=32)
model.fit_partial(
x_train[0],
item_features=dataset_item_features,
user_features=dataset_user_features,
)
"""
as a result we have trained:
model.item_embeddings - matrix of latent vectors with shape (10, 32) - latent vector for each feature
model.user_embeddings - matrix of latent vectors with shape (14, 32)
"""When we predict scores for users and items we make dot product of user_embeddings and item_embeddings. If we have multiple features for one item, we make weighted product of matrixes (see method - model.get_item_representations()) where features - dataset_item_features from my toy example and item_embeddings - latent vectors for each feature.
By the way, each feature represented by latent vector which consist of num_components. if you have your own latent vectors, I think you can "hack" thinks by assign your doc2vec vectors to item_embeddings param of lightfm model (you should be careful with mappings in your dataset, and remember that they are will be changed after training). I don't know will it work or not, but you can try.
I think there is nothing to do with numerical features here as they are not supported here.
HenrikRehbinder
julioasotodv
I'd like to use for example age as a user feature or the components of e.g. a Doc2Vec as item features. Take the doc2Vec as example. Say I have 5 compoments docvec=[1.2, -1.0, 3.4, -4.9, 0.0] fortthe first item and [2.2, ....] for the second and so on. I can create columns so that my item feature sparse matrix is item= [ [0, 0, 1.2], [0, 1, -1.0], [0, 2, 3.4], [0, 3, -4.9] [0,0, 0.0], [1, 0, 2.2], [1, 1, ...] ] The matris is really not sparse, but is anyway represented as such. This works in the sense that fit() gives a result but it takes some time. My question is: Is this supposed to work? the article doesn't say explicitly yes or no. Any help would be appreciated.