ludovikcoba / recoxplainer

recoXplainer

As recommender systems today play an important role in our online experience and are involved in a wide range of decisions, multiple stakeholders demand explanations for the corresponding algorithmic predictions. These demands---together with the benefits of explanations (e.g., trust, efficiency, and sometimes even persuasion)--- have triggered significant interest from researchers in academia and industry.

Nonetheless, to the best of our knowledge, no comprehensive toolkit for explainable recommender systems is available to the community yet. Instead, researchers are frequently faced with the challenge of re-implementing prior algorithms when creating and evaluating new approaches. Aiming to address the resulting need, we introduce RecoXplainer, a software toolkit that includes several state-of-the-art explainability methods, and two evaluation metrics.

Install

Pre-requirements

The following toolkits are necessary:

• conda
• git

Clone and environment set-up

Clone the repo:

git clone https://github.com/ludovikcoba/recoxplainer.git

Create environment on conda:

conda create -n recoxplainer python=3.6 

RecoXplainer was developed with python 3.6. Activate the new environment:

conda activate recoxplainer

Dependencies

Install torch as explained in https://pytorch.org/, we are using the version without CUDA.

When torch is installed navigate to the folder where you cloned the library and run:

pip install -r requirements.txt

This command will install all the dependencies. Next, install the recoxplainer:

pip install -e .

And finally run the notebooks:

jupyter notebook

Running times

Running times for pre-processing, training, recommendation, explanation, and evaluation can be found in the `running_times.cvs' file. The reported time was calculated using a MacBook Pro 2,3 GHz Dual-Core Intel Core i5, 8 GB 2133 MHz LPDDR3.

recoxplainer's Issues

IndexError: single positional indexer is out-of-bounds

I was trying to implement ALS recoxplainer to explain the predictions got by surprise SVD model. But seems like SVD algorithm is building sparse matrix by taking the trainset only. But ALS is building the matrix for entire dataset unique values and filling the trainset values inside. How can i fix this issue while applying for SVD model.