The objective of this project is to provide Data Scientist a framework which
can be easily be extended to rapidly develop recommendation systems. The idea
is to provide a series of core python classes using popular libraries like
Turi Create and
Tensorflow-Recommenders that
provide a proxy for building recommenders using both explicit and implicit
data.
For this project we will be leveraging poetry to
handle installing dependencies and versioning.
To configure your environment, first initialize a virtual environment and then install poetry:
pip install poetry
Next, to install the exact versions and resolving any conflicts, run the following command which poetry reads from the poetry.lock file. This helps ensure everyone is using the same versions of the dependencies.
poetry install
To exclude any of the development dependencies, you can also run:
poetry install --no-dev
Please refer to the series of helper notebooks in the examples/ for examples on incorporating the recommender classes into your own projects. For the demos, feel free to use your own dataset or explore with the movielens dataset as well. Currently, the examples are based off an anonymous online learning platform dataset that contains thousands of user/item interactions. For more details, please see the ".csv' files in data/.
-
CollaborativeFiltering - A simple Collaborative Filtering class based on the following similarity distance metrics such as jaccard, cosine, and pearson. This class can be used to rank similar users or items based on the learned latent features.
-
DeepMatrixFactorization - is an extension of Matrix Factorization for building recommendation engines using explicit data. The modification is to simply apply a non-linear kernel to model the latent feature interactions.
-
RankingFactorizationRecommender - This class builds a Factorization Machine model to learn latent factors for each user and item interaction for implicit data (e.g. implicit Alternating Least Squares). As opposed to performing Matrix Factorization, we can now include slide features that may be numeric or categorical.
-
HybridRecommender - one of the advantage of using neural networks for recommendation systems is the ability to create an architecture that utilizes both the collaborative and content based filtering approaches. This class exploits using explicit data to include side features for user/items. Ideally this type of approach could help address the cold start problem or refined ranking given a subset of items filtered from a candidate generator model (e.g. see retrieval.py)
-
CandidateGeneration - train a two-tower DNN model to produce the top 'k' recommendations by passing through the query tower and finding its representation from the learned embedding vector. An affinity score between the query and all the candidates is calculated and sorted with the k-nearest candidates to the query. The approach implemented is a brute-force method, but it's recommended to implement Approximate Nearest Neighbor (ANN), which is significantly faster.
If interested in deploying your model, a basic Flask example is provided in
sever/ which can be used to make a RESTful api calls to a backend datastore.
This was just a quick prototype, but more to come on this section though and
example on serving a DNN model using Tensorflow-Serving.
If you are interested in contributing ideas on how to make this project better, thoughts on other types of recommenders to include, or to simply report any bugs, please feel free to open up an issue to discuss further. All contributions are welcomed!