[research] deps2vec #100
Description
This is a replacement for #65 which lasted too long to be effective.
Context
Theory
Two papers in MSR 2019 have shown that we can use dependencies graphs to extract information. Namely:
- Identifying Experts in Software Libraries and Frameworks among GitHub Users showed we could profile developers via their modification of the dependency graph.
- Import2vec: Learning Embeddings for Software Libraries showed we could create meaningful embeddings for libraries from large dependencies graph.
Business
As detailed in the ML e-book, there are multiple business use-cases related to the dependency graph:
- Grouping projects by dependency similarity, either in 2D or answering nearest neighbours queries.
- Finding frequent dependency sets, the same way as retailers determine frequent product baskets.
- Finding competing dependencies.
- Suggesting new dependencies.
- Recommending alternatives.
It is also a strong possibility that we could leverage dependencies graph to group similar developers, and assess expertise in libraries or ecosystems.
Objective
We want to use the dependency graph of PGA to create embeddings for as much libraries as possible, and explore how we could use these to answer the business perspectives described above.
Checklist
- Extract the dependencies from the ClickHouse DB - we will do this with some granularity, but it is expected that packages/class imports will simply introduce noise, and will be a pain to extract. We will not care for aliases as well.
- Postprocess the dataset
- Normalize local import
- Link dependencies with different names, keeping in mind that we want to separate versions when possible
- Convert dataset to a sparse representation.
- Create the embeddings with Swivel.
- Per language
- Globally
- Evaluate the quality of the embeddings.
- Qualitatively
- With SGNS-like model, compare with random
- Robustness with K-NN, see previous issue.
- Explore how to create sensible project/developer representations for th eembeddings, and test on orgs.