LightFM how to make predictions for new users (cold start) - user id 8 not in user id mappings

I am building a recommendation system in order to recommend training to employees based on user features and item features which LightFM according to the documentation its a great algorithm.

my user dataframe:

	User-Id	name	age	los	ou	gender	skills
0	1	Luis	21	IFS	architecture	M	python
1	2	Peter	22	ADV	pmo	M	pm
2	3	Jurgen	23	IFS	architecture	M	sql
3	4	Bart	24	IFS	architecture	M	python
4	5	Cristina	25	ADV	pmo	F	pm
5	6	Lambert	33	IFS	development	M	sql
6	7	Rahul	44	IFS	development	M	python

My trainingds data frame

Training-Id	training name	main skill
0	1	basic python	python
1	2	advanced python	python
2	3	basic scrum	pm
3	4	advanced scrum	pm
4	5	basic sql	sql
5	6	advanced sql	sql

My training taken dataframe (10 means a user took that training)
so my weights its only 10s

User-Id	Training-Id	TrainingTaken
0	1	1	10
1	1	2	10
2	2	3	10
3	2	4	10
4	3	5	10
5	3	6	10
6	4	1	10
7	4	2	10

I found this great helpder in order to create tha matrices:
https://github.com/Med-ELOMARI/LightFM-Dataset-Helper

So:

items_column = "Training-Id"
user_column = "User-Id"
ratings_column = "TrainingTaken"

items_feature_columns = [
    "training name",
    "main skill"
]

user_features_columns = ["name","age","los","ou", "gender", "skills"]

dataset_helper_instance = DatasetHelper(
    users_dataframe=usersdf,
    items_dataframe=trainingsdf,
    interactions_dataframe=trainingstakendf,
    item_id_column=items_column,
    items_feature_columns=items_feature_columns,
    user_id_column=user_column,
    user_features_columns=user_features_columns,
    interaction_column=ratings_column,
    clean_unknown_interactions=True,
)

# run the routine
# you can alslo run the steps separately one by one | routine function is simplifying the flow
dataset_helper_instance.routine()

the above helper returns the interaction matrix, the weight matrix, etc.

dataset_helper_instance.weights.todense()
Output menu
matrix([[10., 10.,  0.,  0.,  0.,  0.],
        [ 0.,  0., 10., 10.,  0.,  0.],
        [ 0.,  0.,  0.,  0., 10., 10.],
        [10., 10.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.,  0.]], dtype=float32)


dataset_helper_instance.interactions.todense()

matrix([[1., 1., 0., 0., 0., 0.],
        [0., 0., 1., 1., 0., 0.],
        [0., 0., 0., 0., 1., 1.],
        [1., 1., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.],
        [0., 0., 0., 0., 0., 0.]], dtype=float32)

Then I train test split and fit the model

from lightfm import LightFM
from lightfm.cross_validation import random_train_test_split
(train, test) = random_train_test_split(interactions=dataset_helper_instance.interactions, test_percentage=0.2)

model = LightFM(loss='warp')

model.fit(
    interactions=dataset_helper_instance.interactions,
    sample_weight=dataset_helper_instance.weights,
    item_features=dataset_helper_instance.item_features_list,
    user_features=dataset_helper_instance.user_features_list,
    verbose=True,
    epochs=50,
    num_threads=20,
)

Then I check the AUC and precision:

from lightfm.evaluation import precision_at_k
from lightfm.evaluation import auc_score

train_precision = precision_at_k(model, train,item_features=dataset_helper_instance.item_features_list, user_features=dataset_helper_instance.user_features_list , k=10).mean()
test_precision = precision_at_k(model, test, item_features=dataset_helper_instance.item_features_list, user_features=dataset_helper_instance.user_features_list,k=10).mean()
train_auc = auc_score(model, train,item_features=dataset_helper_instance.item_features_list, user_features=dataset_helper_instance.user_features_list).mean()
test_auc = auc_score(model, test,item_features=dataset_helper_instance.item_features_list, user_features=dataset_helper_instance.user_features_list).mean()
print('Precision: train %.2f, test %.2f. '% (train_precision, test_precision))
print('AUC: train %.2f, test %.2f.' % (train_auc, test_auc))

Precision: train 0.15, test 0.10. 
AUC: train 0.90, test 1.00.

Then I do predictions for an existing user

scores = model.predict(user_ids=6, item_ids=[1,2,3,5,6])
print(scores)

[ 0.01860116 -0.20987387  0.06134995  0.08332028  0.13678455]

Great, I can get some predictions of trainings to follow for user ID 6.

Now I want to predict for new users, (cold start)

I tried the following:

dataset = Dataset()
new_user_feature = [8,{'name:John', 'Age:33', 'los:IFS','ou:development', 'skills:sql'} ]    
new_user_feature = [8,new_user_feature]
new_user_feature = dataset.build_user_features([new_user_feature])
#predict new users User-Id	name	age	los	ou	gender	skills
model.predict(0, item_ids=[1,2,3,5,6], user_features=new_user_feature)

However I get this error:

ValueError: user id 8 not in user id mappings.

What am I missing here?

Start with the Dataset(), build_user_features(), and model.predict() entry points shown in the report, then trace how user IDs and user features are mapped. Reproduce the prediction attempt with the supplied seven-user data and the new user id 8. Done means establishing whether the cold-start call is supported and documenting the correct behavior or limitation.