dandxy89/bellkoralgorithm

A Python implementation of the Bellkor Algorithm

Project	Created	Updated	Version
Re-implementing the Bellkor Algorithm	06/01/2018	05/04/2020	0.1.0

FYI - This is no-longer actively maintained!

Motivation

Short Answer

Implement an Algorithm that had previously won a Data Science / well known competition, in this case the Netflix Grand Prize.

Long Answer

Firstly, as companies increasingly adopt data-driven approaches and prioritize enhancing customer experiences, recommendation systems are gaining prominence. It's evident that personalization is crucial for fostering customer loyalty and engagement. Therefore, delving into the intricacies and implementation of these algorithms seems opportune for advancing my career.

Secondly, I relish the task of building algorithms and systems from scratch in my spare time, primarily to master the art of translating intricate concepts into code. Moreover, the paper's elucidation of feature engineering used for rating prediction is exceptionally clear and succinct. Lastly, I find great value in the challenge of retracing the footsteps of original researchers and developers to understand their implementation strategies, which aligns with my preferred learning approach.

Bellkor Algorithm

This is an implementation of the “BellKor’s Pragmatic Chaos” final solution, which eventually went on to win the Netflix Grand Prize.

The purpose of this algorithm is to predict the ratings using the Bellkor Algorithm. The algorithm uses novel Feature Engineering in order to model the temporal properties of the users and times over time.

Demo and usage

I've created a IPython Notebook here which showcases the implementation.

Acknowledgments

'Yehuda Koren and the Bellkor Pramatic Choas Team' - for documenting the solution in the paper, located here.

Links

• MovieLens 20M dataset - differs from the original but same as the one used in the demo.
• Netflix Grand Prize Info - overview of the original competition.

Testing

Run the following command to start the test suite:

    pytest tests

Coverage

    67%

TODO

• Document results using the MovieLens 20M Dataset
• 2 Part - Blog Post (1 of 2 complete.)

bellkoralgorithm's Issues

calculation of gradients maybe wrong

In Algorithm.py, get_gradients function, your code is shown as follow:

gradients = Gradients(b_u=-2 * error + 2 * self.REGULARISATION.b_u * rs.b_u,
alpha_u=2 * error * (-rs.Dev + 2 * self.REGULARISATION.alpha_u * rs.alpha_u),
b_ut=2 * error * (-1 + self.REGULARISATION.b_ut * rs.b_ut),
b_i=2 * error * (-rs.c_u - rs.c_ut + 2 * self.REGULARISATION.b_i * rs.b_i),
b_ibin=2 * error * (-rs.c_u - rs.c_ut + 2 * self.REGULARISATION.b_ibin * rs.b_ibin),
c_u=2 * error * (-rs.b_i - rs.b_ibin + 2 * self.REGULARISATION.c_u * (rs.c_u - 1)),
c_ut=2 * error * (-rs.b_i - rs.b_ibin + 2 * self.REGULARISATION.c_ut * rs.c_ut),
p=2 * error * (rs.q + 2 * self.REGULARISATION.p * rs.p),
q=2 * error * (rs.p + rs.alpha_p * rs.Dev + 2 * self.REGULARISATION.q * rs.q),
alpha_p=2 * error * (rs.q * rs.Dev + self.REGULARISATION.alpha_p * rs.alpha_p))
——————————————————————————————————————

here's my version, which I think is correct.
+++++++++++++++++++++++++++++++++++++
gradients = Gradients(b_u=-2 * error + 2 * self.REGULARISATION.b_u * rs.b_u,
alpha_u=2 * error * (-rs.Dev) + 2 * self.REGULARISATION.alpha_u * rs.alpha_u,
b_ut=2 * error * (-1) + self.REGULARISATION.b_ut * rs.b_ut,
b_i=2 * error * (-rs.c_u - rs.c_ut) + 2 * self.REGULARISATION.b_i * rs.b_i,
b_ibin=2 * error * (-rs.c_u - rs.c_ut) + 2 * self.REGULARISATION.b_ibin * rs.b_ibin,
c_u=2 * error * (-rs.b_i - rs.b_ibin) + 2 * self.REGULARISATION.c_u * (rs.c_u - 1),
c_ut=2 * error * (-rs.b_i - rs.b_ibin) + 2 * self.REGULARISATION.c_ut * rs.c_ut,
p=2 * error * (-rs.q) + 2 * self.REGULARISATION.p * rs.p,
q=2 * error * (-rs.p - rs.alpha_p * rs.Dev) + 2 * self.REGULARISATION.q * rs.q,
alpha_p=2 * error * (-rs.q * rs.Dev) + self.REGULARISATION.alpha_p * rs.alpha_p)

Please check it. btw, this project is really impressive, thank you very much:)

wrong gradient

hello @dandxy89
I think
b_u = 2 * error * (-1 + 2 * self.REGULARISATION.b_u * rs.b_u)
is wrong
and
b_u = -2 * error + 2 * self.REGULARISATION.b_u * rs.b_u
is correct