Ir para o conteúdo

Software livre Brasil

 Voltar a Blog
Tela cheia

Contributing with Debian Recommendation System

29 de Julho de 2016, 13:58 , por Luciano Prestes Cavalcanti - 0sem comentários ainda | Ninguém está seguindo este artigo ainda.
Visualizado 1173 vezes
Hi, my name is Luciano Prestes, I am participating in the program Google Summer of Code (GSoC), my mentor is Antonio Terceiro, and my co-mentor is Tassia Camoes, both are Debian Developers. The project that I am contributing is the AppRecommender, which is a package recommender for Debian systems, my goal is to add a new strategy of recommendation to AppRecommender, to make it recommend packages after the user installs a new package with 'apt'.
At principle AppRecommender has three recommendation strategies, being them, content-based, collaborative and hybrid. To my work on GSoC this text explains two of these strategies, content-based and collaborative. Content-based strategy get the user packages and analyzes yours descriptions to find another Debian packages that they are similar to the user packages, so AppRecommender uses the content of user packages to recommender similar packages to user. The collaborative strategy compare the user packages with the packages of another users, and then recommends packages that users with similar profile have, where a profile of user is your packages. On her work, Tassia Camoes uses the popularity-contest data to compare the users profiles on the collaborative strategy, the popularity-contest is an application that get the users packages into a submission and send to the popularity-contest server and generates statistical data analyzing the users packages.
I have been working with a classmate on our bachelor thesis since August 2015, in our work we created new strategies to AppRecommender, one using machine-learning and another using a deterministic method to generates the recommendation, another feature that we implemented its improve the user profile using the recently used packages to makes the profile. During our work we study the collaborative strategy and analyzed that strategy and remove it from AppRecommender, because this implementation of collaborative strategy needs to get the popularity-contest submissions on the user's pc, and this is against the privacy policy of popularity-contest.
My work on Google Summer of Code is create a new strategy on AppRecommender, as described above, where this strategy should be able to get an referenced package, or a list of referenced packages, then analyze the users packages making a recommendation using the referenced packages such as base, example: if users run "$ sudo apt install vim", the AppRecommender use "vim" as referenced package, and should recommender packages with relation between "vim" and the other packages that user has installed. This new strategy can be implemented like a content-based strategy, or the collaborative strategy.
The first month of Google Summer of Code its destined to students knows the community of the project, so I talk with the Debian community about my project, to get feedback and ideas about the project. I talk with Debian community on IRC channels, and then came the idea to use the data of popularity-contest to improve the recommendations. Talking with my mentors, they approve the idea of usage popularity-contest data, so we started a discussion about how to use the popularity-contest data on AppRecommender without broken the privacy policy of popularity-contest.
Now my work on Google Summer of Code is create the new strategy for AppRecommender that can makes recommendation using a list of packages as reference, so as explained above, when user install packages like "sudo apt install vim vagrant", AppRecommender should recommends packages with relation between the packages "vim" and "vagrant", and this recommendation should be relation with the user profile. The other work its use the popularity-contest data to improve the recommendations of AppRecommender using a new model of collaborative strategies.

0sem comentários ainda

Enviar um comentário

Os campos são obrigatórios.

Se você é um usuário registrado, pode se identificar e ser reconhecido automaticamente.