This repository provides a reference implementation of vec2vec, which can reduce the dimension to matrix.
For the algorithm of vec2vec, you can read the following paper:
Xiang Wang, Xiaoyong Li, Junxing Zhu, et al.
A Local Similarity-Preserving Framework for Nonlinear Dimensionality Reduction with Neural Networks[C].
International Conference on Database Systems for Advanced Applications (Dasfaa 2021).
Springer, Cham, 2021: 376-391.
Before starting this project, you must install requirements below.
faiss==1.7.0
gensim==4.0.1
networkx==2.6.2
scikit-learn==0.24.2
Note：It's recommended that using conda to install faiss, and conda version need to update.
conda install faiss-cpu -c conda-forge
-
To run vec2vec by terminal, execute the following command from the project home directory:
python ./vec2vec/main.py --input ./vec2vec/data/train.bowYou can check out the other options available by using:
python ./vec2vec/main.py --help -
To run the vec2vec in your project, execute the following command:
pip install vec2vec
Refer to the ./vec2vec/data/train.bow in the project.
The output are like below:
************* The number of num_walks is : 5 *******************
Matrix2vec p and q and topk: 1 1 10
The shape of the input matrix: (2000, 13155)
BuildNNGraphFromFAISS Finished in 0:00:03.305026 s.
The shape of the adjmatrix is: (2000, 2000)
Preprocess_transition_probs Finished in 0:00:00.902988 s.
Random Walk Finished in 0:00:00.590795 s.
Begin to train word2vec...
Model Matrix2vec Finished in 0:00:09.073013 s.
Accuracy： [0.662 0.652 0.64 0.668]
Accuracy: 0.6555 (+/- 0.0212)
Please send any questions you might have about the code and/or the algorithm to xiangwangcn@nudt.edu.cn.