swapuniba/clayrs

Complexly represent contents, build recommender systems, evaluate them. All in one place!

clayrs's Issues

[Question/Request] Is it possible to make the framework back-end agnostic?

Is your feature request related to a problem? Please describe.
Currently, we have an interesting variety of possible back-ends for mathematical processing. However, to minimize the differences and allow a quick transition between them, frameworks like Ivy are growing and allowing users/developers/researchers/etc to implement their models through an agnostic interface and execute them in any back-end they want. This solution would also make PyTorch not a mandatory dependency for the project, and we would be able to install ClayRS along a specific back-end with commands like pip install clayrs[tf] or pip install clayrs[torch] - which also complies with HuggingFace, for instance.

Describe the solution you'd like
Something like this would be interesting if I have more than one backend available:

from clayrs.backend import use_tensorflow
use_tensorflow()

If there is only one backend available, use it as default with an info/warning message in the first usage.

Additional context
Filling this issue while attending Marco Polignano's lecture at RecSys Summer School 2023 :)

[Question/Request] Collaborative Filtering?

Is your feature request related to a problem? Please describe.
Is it a plan to evolve and include Collaborative Filtering in the package? It would be exciting if we could have methods like matrix factorization off-the-shelf, considering that the validation and evaluation are the same in many cases.

Describe the solution you'd like
I want to be able to import a Matrix Factorization with something like this:

from clayrs import recsys as rs
strategy = rs.MatrixFactorization(...)
collaborative_rs = rs.CollaborativeFilteringRS(strategy, ...)

Describe alternatives you've considered
There are a few models already implemented in other frameworks shared during previous editions of RecSys, including DRecPy, which also provides efficient memory-disk management.

Additional context
Filling this issue while attending Marco Polignano's lecture at RecSys Summer School 2023 :)

Add missing Bootstrap Partitioning technique to documentation

Set default number of cpus to 1 for the RecSys module

At the moment, the default number of cpus for the recsys module is 0, meaning that all possible cpus will be used: this is a problem since python multiprocessing relies on pickling contexts which may be slow depending on the environment (mainly colab).

Change the default value to 1 to improve ease of use and speed for the above cases

[Question] Are all the dependencies really mandatory?

Is your feature request related to a problem? Please describe.
Currently, it takes a long time to install the package and it also requires the installation of other resources - such as MySQL - which may often not be used. Is it possible to reduce the mandatory dependencies to the core ones, and require other ones as long as the user needs them?

Describe the solution you'd like
For example, NLTK only downloads text corpora when requested by the user. Following these optics, I am a user that will likely never use such a package for knowledge-aware recommenders or building knowledge graphs, but I have a great interest in using it for Collaborative Filtering - and perhaps, in the future, Sequential and Session-based recommendation.

Additional context
Filling this issue while attending Marco Polignano's lecture at RecSys Summer School 2023 :)

Improve SPARQL query speed for DBPedia

At the moment, retrieving properties from DBPedia is quite slow (20 min for 1682 contents).

Consider the possibility to improve it

Improve coherence Transformer embedding technique

At the moment, the class Transformer of the content analyzer has embedded inside the pooling operation which is instead done for the other embedding techniques by the Words2Sentence class, Words2Doc class, etc.

Fix this to improve coherence