Project within the subject Machine Learning Engineering (IUM). The scope of this project is to create a recommendation system for the fictional company named "eSzoppping". The recommendation system should benefit the company marketing and sales strategy.
In the following project we play the role of the analysts working in the company
"eSzoppping" - an online store with electronics and computer games.
Working on this position is not easy.
The task, which we recently obtained, was formed only as an enigmatic description.
It is up to us to specify the details of the task so that they can be implemented.
This, of course, requires understanding of the problem, analyzing the data
and (sometimes) negotiations with the boss.
Apart from analyzing the issue and training the models,
we must prepare models for production release.
We should assume, that in the future there will be new versions of the product.
Some people come to our website and can't decide which products to take a closer look at. Maybe we could give them some recommendation?
The problem analysis can be found in notebook called "problem_analysis.ipynb". Unfortunately, it is available only in Polish version.
The data analysis can be found in notebook called "data_analysis.ipynb". Unfortunately, it is available only in Polish version.
As the solution to the problem, we have implemented 2 models serving the recommendations
to the users.
Model number 1 is a simple model, serving recommendations
based on products metric score. Metric used is similar to the one used by IMDB
recommendation systems.
Model number 2 is more advanced model. It is based on the collaborative filtering and grouping method. It groups users based on their interests and for each group it serves most popular products within this group.
Models have been tested with the usage of custom testing method.
The method is based on session splitting and predicting the
outcomes of next user activity. Metric used for testing purposes
is the accuracy.
Model 1 results: 40% of accuracy.
Model 2 results: 74% of accuracy.
Models can be released to production with the usage of prepared micro-service. Micro-service implementation uses Flask-restful framework. The service API is available in the JSON form.
The final documentation of the project contains many more useful and exhausting side-notes. Unfortunately, it is available only in Polish version.