the-algorithm-ml/projects/twhin
2023-09-14 11:30:10 +05:30
..
config Twitter's Recommendation Algorithm - Heavy Ranker and TwHIN embeddings 2023-03-31 13:05:14 -05:00
data update 2023-09-14 11:30:10 +05:30
models update 2023-09-14 11:30:10 +05:30
scripts Twitter's Recommendation Algorithm - Heavy Ranker and TwHIN embeddings 2023-03-31 13:05:14 -05:00
config.py Twitter's Recommendation Algorithm - Heavy Ranker and TwHIN embeddings 2023-03-31 13:05:14 -05:00
machines.yaml Twitter's Recommendation Algorithm - Heavy Ranker and TwHIN embeddings 2023-03-31 13:05:14 -05:00
metrics.py Twitter's Recommendation Algorithm - Heavy Ranker and TwHIN embeddings 2023-03-31 13:05:14 -05:00
optimizer.py update 2023-09-13 11:22:13 +05:30
README.md Twitter's Recommendation Algorithm - Heavy Ranker and TwHIN embeddings 2023-03-31 13:05:14 -05:00
run.py update 2023-09-13 11:22:13 +05:30
test_optimizer.py Twitter's Recommendation Algorithm - Heavy Ranker and TwHIN embeddings 2023-03-31 13:05:14 -05:00

Twhin in torchrec

This project contains code for pretraining dense vector embedding features for Twitter entities. Within Twitter, these embeddings are used for candidate retrieval and as model features in a variety of recommender system models.

We obtain entity embeddings based on a variety of graph data within Twitter such as: "User follows User" "User favorites Tweet" "User clicks Advertisement"

While we cannot release the graph data used to train TwHIN embeddings due to privacy restrictions, heavily subsampled, anonymized open-sourced graph data can used: https://huggingface.co/datasets/Twitter/TwitterFollowGraph https://huggingface.co/datasets/Twitter/TwitterFaveGraph

The code expects parquet files with three columns: lhs, rel, rhs that refer to the vocab index of the left-hand-side node, relation type, and right-hand-side node of each edge in a graph respectively.

The location of the data must be specified in the configuration yaml files in projects/twhin/configs.

Workflow

  • Build local development images ./scripts/build_images.sh
  • Run with ./scripts/docker_run.sh
  • Iterate in image with ./scripts/idocker.sh
  • Run tests with ./scripts/docker_test.sh