Skip-Grammar

A framework for representing sequences as embeddings.

Models

Skip-gram Negative Sampling (SGNS)

Popular natural language processing models such as word2vec and bert can be repurposed to learn relationships from arbitrary sequences of items. Skip-gram Negative Sampling is such an algorithm part of the models module. This is implemented in PyTorch components or can be composed as a PyTorch Lightning module. Both are availble under the relevent namespaces skipgrammar.models.sgns and skipgrammar.models.lighting.sgns.

Datasets

Last.FM

The Last.FM Dataset-1K dataset is comprised of the listening history of approximately 1,000 users from the music service Last.FM. The dataset is availble at the project's main site here and also preprocessed here for ease of use. The variants in the dataset module use the latter.

MovieLens

The popular recommendation system dataset MovieLens is availble in three variants via the dataset module.