Skip to content

Tf-Rec is a python💻 package for building⚒ Recommender Systems. It is built on top of Keras and Tensorflow 2 to utilize GPU Acceleration during training.

License

Notifications You must be signed in to change notification settings

Praful932/Tf-Rec

Repository files navigation

Tf-rec

Tf-Rec is a python💻 package for building⚒ Recommender Systems. It is built on top of Keras and Tensorflow 2 to utilize GPU Acceleration during training.

PyPI version Python contributions welcome
Tests Publish(Package) Deploy(Docs)
Hits GitHub stars GitHub forks GitHub watchers

Contents

Why Tf-Rec? 🧐

There are several open source libraries which implement popular recommender algorithms in, infact this library is inspired by them - Surprise and Funk-SVD. However, there is bottleneck in training time, when the training data is huge. This can be solved by using ready frameworks like Tensorflow 2 & Keras which support running computations on GPU thus delivering speed and higher throughput. Building on top of such frameworks also provide us with off the shelf capabilities such as using different optimizers, Data API, exporting the model to other platforms and much more. Tfrec provides ready implementations of algorithms which can be directly used with few lines of Tensorflow Code. Currently this library supports these algorithms.

Installation ⚡

The package is available on PyPi:

pip install tfrec

Quick Start & Documentation 📝

API Docs

SVD Example

from tfrec.models import SVD
from tfrec.datasets import fetch_ml_100k
from tfrec.utils import preprocess_and_split
import numpy as np

data = fetch_ml_100k()
dataset, user_item_encodings = preprocess_and_split(data)

(x_train, y_train), (x_test, y_test) = dataset
(user_to_encoded, encoded_to_user,item_to_encoded, encoded_to_item) = user_item_encodings

num_users = len(np.unique(data['userId']))
num_movies = len(np.unique(data['movieId']))
global_mean = np.mean(data['rating'])

model = SVD(num_users, num_movies, global_mean)
model.compile(loss = 'mean_squared_error', optimizer = 'adam')
model.fit(x_train, y_train)
2521/2521 [==============================] - 11s 4ms/step - loss: 0.9963

SVD++ Example

from tfrec.models import SVDpp

model = SVDpp(num_users, num_movies, global_mean)
# Needs to be called before fitting
model.implicit_feedback(x_train)
model.compile(loss = 'mean_squared_error', optimizer = 'adam')

model.fit(x_train, y_train)
2521/2521 [==============================] - 49s 20ms/step - loss: 1.0332

KFold Cross Validation Example

from tfrec.utils import cross_validate
model = SVD(num_users, num_movies, global_mean)
model.compile(loss = 'mean_squared_error', optimizer = 'adam', metrics=['mae','RootMeanSquaredError'])
all_metrics = cross_validate(model, x_train, y_train)
Mean Loss : 0.899022102355957
Mean Mae : 0.6596329569816589
Mean Root_mean_squared_error : 0.8578477501869202

Supported Algorithms 🎯

Currently the library supports these algorithms:

  • SVD - Despite the Name, it is different from the Eigen Decomposition of Assymmetric Matrices. In a gist, it approximates a vector for each user and each item. The vector contains latent factors which signify for brevity sake, if the item is a movie the movie vector would represent - how much the movie contains action or romance likewise. Similarly for the user. The predicted rating is given by:

  • SVD++ - This is an extension of SVD which incorporates implicit feedback, by also taking into account the interactions between the user and the item by involving another factor. More Precisely, it takes into account the fact that the user has rated an item itself as a preference than an item which the user has not rated. The predicted rating is given by:
    image

Benchmark 🔥

Both of the algorithms were tested on Google Collab using a GPU Runtime. The dataset used was the MovieLens-100k. Default parameters were used for intilization of Model. Optimizer used was Adam and batch size used was 128. These are the 5-Fold Cross Validation Scores:

Algorithm Mean MAE Mean RMSE Time per Epoch
SVD 0.6701 0.8694 < 3 sec
SVD++ 0.6838 0.8862 < 45 sec