Social network use case: Handle user to user interactions (items = users, reciprocal)

Nobody has claimed this yet.

Hi there,

first: **Many thanks for the awesome library!**

I would like to ask if the following use case of "social recommendations" is possible with LightFM (in a more or less efficient way).

## Scenario
Suppose a user `u0` who is befriended with `u1` wants to get new user recommendations.
The recommendation goal is not to recommend new items in the classical sense but to recommend new users.

Since there are other *types* of interactions with varying importance, the interaction matrix is weighted (say friendship equals a weight of 1):

```
(0, 1) 1
(1, 0) 1
```

Because friendship is bidirectional, both interactions are in the matrix, but unidirectional (one-sided) implicit interactions such as "view" should also be possible.
If there are multiple interactions between two users, the interaction weights are summed up (smoothed e.g. by using a `log` function).

Consequently, items are not a different entity and items effectively become users.
Now, features such as `gender` or `age` are also present to combat the cold start problem and fully utilise the efficient hybrid recommendation capabilities of LightFM.

LightFM is trained as follows with 2 effective inputs (`interaction_matrix` and `user_matrix`, both built using the `Dataset` class):

```python
lightfm.fit(interactions=interaction_matrix, sample_weight=interaction_matrix,
item_features=user_matrix, user_features=user_matrix)
```

To save memory, the `interaction_matrix` is reused for the weights, too ([as stated in the docs](https://making.lyst.com/lightfm/docs/lightfm.html)).
Since the items' features (i.e. what is to be recommended) do not differ from the users' features, the same matrix is used.

## Questions

1. Is the above scenario possible? If yes, is there a way to make it more efficient?
2. As far as I understand, the embeddings and biases for users and items are learned separately using multi-threaded SGD. Should the added identity matrix (built by the `Dataset` class) be used for the `item_features` as well as for the `user_features` parameters or is it actually making the model less expressive, preventing transfer learning to occur?

I searched a lot but could not find a similar use case in other issues since most LightFM users want to recommend real `items != users` (movies, products, questions, etc.).

**Any help or suggestions appreciated!**
If LightFM is not suitable for the mentioned scenario, that would be good to know, too.

[Edited title: I've read that such specializations of recommender systems are also called "reciprocal".]

Start with the LightFM documentation linked in the issue and the Dataset and fit entry points described in the example. Compare the interaction_matrix, sample_weight, item_features, and user_features inputs for the reciprocal-user scenario. Done means establishing whether this use case is supported and documenting any efficiency or identity-feature guidance.

Written by the indexing model from the issue text.