This repository consists of the implementation of a unified recommender system framework for easily integrating matrix and tensor factorization methods.
The methods included are:
- Non-negative Matrix Factorization (NMF)
- TransE
- TransH
- TransR
- TransD
- RESCOM, a modification of RESCAL
- NECTR
- [
Python >=3.6] (https://www.python.org/) - [
TensorFlow >=1.9.0] (https://www.tensorflow.org/) - [
pandas >=0.23.4] (https://pandas.pydata.org/) - [
NumPy >=1.15.0] (http://www.numpy.org/) - [
scikit-learn >= 0.19.2] (https://scikit-learn.org/)
The models expect data to be given in the following form:
- triple2id.txt: All triples that make up the knowledge graph in the form entity1 x entity2 x relation12. The first line is the number of triples to be used for training (typically, the total number of lines). This file can also contain data in the form entity1 x entity2 x relation12 x count12 where the last entry represents the count, i.e., the number of occurences of the triple in the knowledge graph. It is especially important for the solutions as the count represents the number of times a specific item has been configured in the corresponding solution. This format is necessary for NECTR as an input to the neural autoencoder.
- entity2id.txt: All entities and the corresponding Ids, one per line. The first line is the number of entities.
- relation2id.txt: All relations and the corresponding Ids, one per line. The first line is the number of relations.
- Categories.csv: Each line has information about a category and the items belonging to that category. This file is used in the category-based evaluation and for visualization.
All text files use \t as the column delimiter and the CSV files use ;.
The automation solution data is available upon request ([email protected]).
A demo on how to use the recommender system can be found in the notebook.