recommendations for cold start users when model trained without user_features?

Nobody has claimed this yet.

Due to the getting the same recommendations for almost all users, I trained the model without user_features as was suggested in [here](https://github.com/lyst/lightfm/issues/320#issuecomment-401329449) (~90 unique items are recommended). The model training is like so:
```
model = LightFM(no_components= embedding_dimension, learning_rate=0.01, loss='bpr',
learning_schedule = 'adadelta', random_state=seed)
model.fit(train_interactions,
epochs=NUM_EPOCHS,
num_threads=20,
verbose=True)
```
During the inference, for the cold and existing customers, I did the following as was suggested [here](https://github.com/lyst/lightfm/issues/210#issuecomment-324387779):
```
if not cold_start_user_flg:
scores = model.predict(user_ids=user_id_map[user_id],
item_ids=np.arange(n_items),
num_threads=20
)
else:
scores = model.predict(user_ids=0,
item_ids=np.arange(n_items),
user_features=user_feature_cold_start,
num_threads=20
)
```
My question is since the model got trained without user_features, how is it even able to provide recommendations for cold-start users?

Start with the linked discussions and the fit and predict calls shown in the report. Trace how user_features are supplied during training and inference, then document whether this usage can provide cold-start recommendations and what inputs are required; done means the behavior is clearly explained.

Written by the indexing model from the issue text.