DeepWalk-implementation

This repository consists of DeepWalk implementations one using gensim word2vec model and second entirely in pytorch. Two implementations are then compared on task of outlier detection on dataset inj_cora coming from https://github.com/pygod-team/data

Files description

Class and function definitions

DeepWalkGensim - consists of DeepWalkGensim class implementation that performs DeepWalk using gensim skipgram implementation

DeepWalkOur - consists of DeepWalkOurs class implementation that performs DeepWalk using our implementation of Huffman tree, Hierarchical Softmax and SkipGram model

PairsDataset - consists of PairsDataset class implementation that is a torch Dataset class that accepts list of pairs of (center node, context node)

HierarchicalSoftmax - consists of HierarchicalSoftmax layer implementation

HuffmanTree - consists of HuffmanNode and HuffmanTree class implementation

SkipGram - consists of NeuralEmbedder class that is SkipGram model implementation with HierarchicalSoftmax class if specified by user

helpers - consists of helpers functions definitions. read_from_pyg(...) read .pt graph file, predict(...) given embeddings predicts outliers using XGBoost or LogisticRegression classifier on 20% test train split, create_freq_dict(...) creates dictionary of number of visits in each node while doing a walk of given lenght from each node.

Scripts

experiment1_gensim_grid_search - script that performs grid search on DeepWalk with gensim SkipGram implementation

experiment2_our_deepwalk_training - script that performs grid search on DeepWalk with our SkipGram implementation

Python notebook files

Graphs2.ipynb - summarization of experiment 1 and creation of sensitivity graphs

Benchmarking.ipynb - benchmark of other models