DeepNeighbor

Embedding-based Retrieval for ANN Search and Recommendations!
View Demo · Report Bug · Request Feature

DeepNeighbor is a High-level,Flexible and Extendible package for embedding-based information retrieval from user-item interaction logs. Just as the name suggested, 'deep' means deep learning models to get user/item embeddings, while 'neighbor' means approximate nearest neighbor search in the embedding space.
It mainly has two parts : Embed step and Search step by the following codes:

model = Embed(data_path); model.train()，which generates embeddings for users and items (Deep),
model.search(), which looks for Approximate nearest neighbor for seed user/item (Neighbor) .

Install

pip install deepneighbor

How To Use

from deepneighbor import Embed

model = Embed(data,model='gat')
model.train()
model.search(seed = 'Louis', k=10)

Input format

The input data for the Embed() should be a (*.csv or *.txt ) file path (e.g. '\data\data.csv')with two columns in order: 'user' and 'item'. For each user, the item are recommended to be ordered by time.

Models & parameters in Embed()

• Word2Vec w2v
• Factorization Machines fm
• Deep Semantic Similarity Model
• Siamese Network with triple loss
• Deepwalk
• Graph convolutional network
• Neural Graph Collaborative Filtering algorithm ngcf
• Matrix factorization mf
• Graph attention network gat

Model Parameters

deepwalk

model = Embed(data, model = 'deepwalk')
model.train(window_size=5,
            workers=1,
            iter=1
            dimensions=128)

• window_size Skip-gram window size.
• workersUse these many worker threads to train the model (=faster training with multicore machines).
• iter Number of iterations (epochs) over the corpus.
• dimensions Dimensions for the node embeddings

graph attention network

model = Embed(data, model = 'gat')
model.train(window_size=5,
            learning_rate=0.01,
            epochs = 10,
            dimensions = 128,
            num_of_walks=80,
            beta=0.5,
            gamma=0.5,)

• window_size Skip-gram window size.
• learning_rate learning rate for optimizing graph attention network
• epochs Number of gradient descent iterations.
• dimensions Dimensions for the embeddings for each node (user/item)
• num_of_walksNumber of random walks.
• beta and gammaRegularization parameter.

How To Search

model.search(seed, k)

• seed The Driver for the algorithms
• k Number of Nearest Neighbors.

Examples

Open Colab to run the example with facebook data.

Contact

Please contact louiswang524@gmail.com for collaboration or providing feedbacks.

License

This project is under MIT License, please see here for details.