Sapling Similarity Collaborative Filtering
This repository contains the implementation for our paper:
Sapling Similarity: a performing and interpretable memory-based tool for recommendation https://arxiv.org/abs/2210.07039
where we introduce Sapling Similarity and SSCF.
Python version and libraries
The python and library versions used to test this code are:
- python 3.8.8
- numpy 1.19.5
- pandas 1.2.4
- sklearn 1.0.2
An example to run the code
To run the code (in the code folder) using sapling similarity on export data with gamma = 0.7:
python main.py --dataset="export" --test="test" --gamma=0.7 --similarity="sapling_similarity"
We provide five processed datasets:
- export
- amazon-product
- gowalla
- yelp2018
- amazon-book
if test="validation", a validation set is used to evaluate the performance. The validation set we used for the results we show in the paper is located in the corresponding data folder. To generate a new validation set you can run the generate_validation.py code.
gamma is the only parameter in our model and its value ranges from 0 (user-based approach) to 1 (item-based approach).
similarity can be one of the following:
- common_neighbors
- jaccard
- adamic_adar
- resource_allocation
- cosine_similarity
- sorensen
- hub_depressed_index
- hub-promoted_index
- taxonomy_network
- probabilistic_spreading
- pearson
- sapling_similarity
Computational time
To test our code we used a computer with the following specifications:
- RAM: 32 GiB
- processor: Intel® Core™ i7-10700 CPU @ 2.90GHz × 16
with our computer the time required to run the SSCF method is:
- export data: 1.13s
- amazon-product: 6.92s
- gowalla: 512.03s
- yelp2018: 512.90s
- amazon-book 11938.65s
Code for Amazon-book data
Our proposed collaborative filtering model SSCF requires the measure of 5 matrices: a similarity matrix for the users, a similarity matrix for the items, and three recommendation matrices: one for the user-based approach, one for the item-based approach and one for the final SSCF model. If we deal with big datasets, depending on the specifications of the computer, it could be not possible to work with these matrices. For instance our computer is not able to run the main.py code when using the amazon-book dataset. We implemented the version main_light.py in which matrices computation is divided in 10 steps (the number of steps may be chosen through the M1 variable in the code) and instead of numpy.array we use numpy.memmap. In each step we compute only a fraction of the columns (or of the rows) of the matrices. using main_light.py our computer can run the SSCF model on amazon-book data.
python main_for_big_dataset.py --dataset="amazon-book" --test="test" --gamma=0.9
The use of np.memmap on Amazon-Book data requires an available disk space of 40 GiB.
Rating prediction
We also provide an implementation to predict ratings with movielens data (in the movielens_rating folder)