chrisvittal / quackin

A recommender systems framework for Rust

Quackin

Release the quackin! 🦆

Quackin is a recommender systems framework written in Rust. This is a young project, which means two things:

• There will be a lot of breaking changes until version 1.0
• Is a perfect project for your contributions ;)

Quackin was concieved as a framework for collaborative filtering recommender systems. But it is open to new ideas. If you want to contribute in some way open or start working on an issue or ping me at [e-mail].

One of the difficulties for developing this crate is that the Rust's machine learning and data mining ecosystem is on an early stage, then there is no standard libraries for anything yet. Nevertheless, I believe that we should take advantage of the control and speed that Rust offers to provide a competitive alternative to frameworks written in languages with a more developed ecosystem like Python or Java.

Getting started

We will write a simple movie recommender engine using Quackin over one of the movielens dataset. Download ml-latest-small.zip from here, and then extract the ratings.csv file.

Loading data into Quackin

First we need to load the movielens dataset into Rust. This dataset has the following columns: userId, movieId, rating and timestamp. We must tell quackin that this file has this columns in this specific order. Also, we need to specify if this file has headers or not and which delimiter is separating the values`.

use quackin::data::{Record, ReadOptions, read_custom_records};
quackin::data::Field::*;

let options = ReadOptions::custom(vec![UserID, ItemID, Rating, Other], true, ',').unwrap();
//                                                             ^^^^^   ^^^^  ^^^
//                                                             |       |     |
//                             we don't care about the timestamp.      |     |
//                                                 this file has headers.    |
//                                                    use comma as a delimiter.

Whew, thats a lot of boilerplate for just loading the dataset. But this was the hardest part of the process. Now lets load the dataset:

let records = read_custom_records("/path/to/movielens", options);

Thats it! now let's build a recommender

Creating a recommender

We will use a traditional K-nearest neighbors algorithm to build a basic recommender. For each user we will take the 50 nearest users to him, using a cosine similarity

use quackin::recommender::{KnnUserRecommender, Recommender};
use quackin::metrics::similarity::cosine;

let recommender = KnnUserRecommender::from_records(&records, cosine, 50);

Now we can ask predictions from the recommender. For example, the user with user ID 1 gave a rating of 4 to the classic science fiction movie "Tron". What would our recommender predict about this?

println!("{:?}", recommender.predict("1", "2105"));
// Ok(3.504942020280084)

Not bad! we aren't that far from the real rating. Now you can start recommending movies to everyone!