reidkr / venmo-word2vec

Word2Vec embeddings: pre-trained using 400M+ Venmo transactions

Simple Word2Vec web service, model pre-trained on Venmo dataset

API serves Word2Vec model pre-trained on about 400M public Venmo transactions posted between 2014 and 2017. The model consists 200-dimensional vectors for 100,000+ words and emojis used on the payment platform. Word2Vec maps words to vectors in such a way that words that have similar meanings or that occur in similar contexts (based on their usage patterns) have vectors that are close together.

Getting started

w/ Pipenv

This project uses Pipenv to manage dependencies (described in the Pipfile) and Python virtual environments.

To get started, install Pipenv with pip:

$ pip install -U pipenv

On macOS, pipenv can also be installed via Homebrew:

$ brew install pipenv

To install the projects' dependencies, in the project's root folder run:

$ pipenv install --dev

Then run the web service:

pipenv run python -m flask run

Usage

Word2Vec automatically learns patterns and relationships in user interactions and the transactions they make.

To interact with the model, visit the http://localhost:5000/w2v_venmo/api/v0.1/most_similar endpoint. It accepts a list of words that contribute positively or negatively to a similarity measure between word vectors and returns the top-N most similar words.

$ curl http://localhost:5000/w2v_venmo/api/v0.1/most_similar \
    --request POST \
    --header "Content-Type: application/json" \
    --data '{"positive": [""], "negative": [""], "topn": 1}'

Examples

# The word vectors encode associations that reflect the usage patterns of emojis. 
# For instance, users most often use the 'eggplant' and 'peach' emojis in the same context.
curl http://localhost:5000/w2v_venmo/api/v0.1/most_similar \
    --request POST \
    --header "Content-Type: application/json" \
    --data '{"positive": ["🍆"]}'

''' 
Out:
"similar": [("🍑", 0.777)]
'''

# Word2Vec also learns slang for certain transactions.
curl http://localhost:5000/w2v_venmo/api/v0.1/most_similar \
    --request POST \
    --header "Content-Type: application/json" \
    --data '{"positive": ["marijuana"]}'

''' 
Out:
"similar": [("weed", 0.751), ("kush", 0.621)]
'''

# The model also infers concepts related to combinations of words 
curl http://localhost:5000/w2v_venmo/api/v0.1/most_similar \
    --request POST \
    --header "Content-Type: application/json" \
    --data '{"positive": ["pizza", "football", "party"]}'

''' 
Out:
"similar": [("superbowl", 0.489)]
'''