MANELA

This repo provides an reference implementation of MANELA simulator. For detailed informations, see.

Organization

This folder is organized as follows.

• baseline:                    Implementations of baseline methods, including node2vec and relationalneighbors
• examples:                  Example datasets.
• __main__.py:              The entrance script for training MANELA embeddings.
• distributed.py:           The main implementation module for MANELA simulator.
• edge_prediction.py:   The module for validating link prediction performance.
• evaluation_utils.py:    A module providing necessary utilities for evaluations.
• graph.py:                   Implementation of networks and their corresponding operations.
• scoring.py:                 A module for validating node classification performance.
• visualize.py:               A module implemented network visualizations.

Usage

Basic: Training MANELA embeddings.

A simple example to run __main__.py is as follows. It takes BlogCatalog network as input file, trains embeddings using MANELA algorithm, and save blogcatalog.embeddings as output file. The input file is assumed to be in .mat format, whereas output files follow the format in Word2vec model of gensim.

python __main__.py --path examples\datasets\blogcatalog.mat --output examples\embeddings\blogcatalog.embeddings --updates 60 --ratio 0.3 --window 10

The optional arguments specify hyperparameters of MANELA algorithm:

• --window: The window size hyperparameter, denoted as w in the paper.
• --ratio: The r₁ component of the ratio vector, where we assume ratio vector is two dimensional
• --updates: The update hyperparameter unique in MANELA simulator, where definitions are: in one unit update, each node v performs d(v)/w iterations. d(v) is the degree of node v. A reasonable rounding mechanism is performed here to guarantee interger iteration numbers.
• --dimension: The dimension of embedding vectors, denoted as d in the paper. Set to 128 by default.
• --negative: The number of negative samples in between two positive updates, denoted as k in the paper. Set to 5 by default.
• --alpha: The initial learning rate. Set to 0.025 by default.
• --seed: Random seed, 0 by default.

Validations: Testing performance in various tasks.

Node Classification

To score the node classification, run:

python scoring.py --emb examples\embeddings\blogcatalog.embeddings --network examples\datasets\blogcatalog.mat --result examples\results\blogcatalog.txt --all --num-shuffles 10

The arguments of scoring module:

• --emb: The embedding path.
• --network: The path of the network dataset.
• --num-shuffles: The number of runs that f1 scores are averaged over.
• --result: The path of the report file. By default it will only display results on the screen without saving them to files.

Link Prediction

A typical command to run link prediction on MANELA:

python edge_prediction.py --input examples\datasets\Homo_sapiens.mat --output examples\results\PPI_link_pred.txt --name manela --v 192 --u 0.4 --round 10 

For comparison purpose, a more convenient way is to run manela, DeepWalk[1] and node2vec[2] on the same network at one time:

python edge_prediction.py --input examples\datasets\Homo_sapiens.mat --output examples\results\PPI_all_link_pred.txt --all --round 10

The arguments of this module:

• --input: The path of the input network.
• --output: The result file path.
• --name: The name of the algorithms being used. Can be toggled between [manela, DeepWalk, node2vec].
• --round: The number of runs. Note that we don't provide averged MAP or Precisions over these runs, but report raw data to you.
• --v and --u: The updates, ratio hyperparameters for manela or p, q hyperparameters for node2vec respectively.
• --all: When this flag is switched on, ignore '--name' option and execute validations on all three algorithms. The corresponding hyperparameters are preset to be optimal as described in the paper.

Visualization

The script is invoked directly:

python visualize.py --ratio 0.75

where --ratio is the r₁ hyperparameter of MANELA.

Requirements

• To train MANELA embeddings: Python3 + Numpy + Scipy
• To run node classification scripts (scoring.py): scikit-learn
• To run link prediction scripts (edge_prediction.py) : scikit-learn + networkx < 2.0
• To run visualization (visualize.py): GEM[3] + matplotlib < 3.0 + networkx < 2.0

Future Works

Build a C extension to run MANELA embeddings training.