recoreco

Fast item-to-item recommendations on the command line.


Keywords
item-collaborative-filtering, recommender-system, rust
License
GPL-3.0-or-later

Documentation

recoreco

Fast item-to-item recommendations on the command line.

GitHub license GitHub issues Build Status

Installation

Currently, the only convenient way to install recoreco is via Rust's package manager cargo:

$ cargo install recoreco

Quickstart

Recoreco computes highly associated pairs of items (in the sense of 'people who are interested in X are also interested in Y') from interactions between users and items.

It is a command line tool that expects a CSV file as input, where each line denotes an interaction between a user and an item and consists of a user identifier and an item identifier separated by a tab character. Recoreco by default outputs 10 associated items per item (with no particular ranking) in JSON format.

If you would like to learn a bit more about the math behind the approach that recoreco is built on, checkout the book on practical machine learning: innovations in recommendation and the talk on real-time puppies and ponies from my friend Ted Dunning.

Example: Finding related music artists with recoreco

As an example, we will compute related artists from a music dataset crawled from last.fm. The data contains 17,535,655 interactions between 358,868 users and 292,365 bands.

As a first step, we download the data, uncompress it and have a look at the format:
$ wget http://mtg.upf.edu/static/datasets/last.fm/lastfm-dataset-360K.tar.gz
$ tar xvfz lastfm-dataset-360K.tar.gz

$ head lastfm-dataset-360K/usersha1-artmbid-artname-plays.tsv
00000c289a1829a808ac09c00daf10bc3c4e223b	3bd73256-3905-4f3a-97e2-8b341527f805	betty blowtorch	2137
00000c289a1829a808ac09c00daf10bc3c4e223b	f2fb0ff0-5679-42ec-a55c-15109ce6e320	die Ärzte	1099
00000c289a1829a808ac09c00daf10bc3c4e223b	b3ae82c2-e60b-4551-a76d-6620f1b456aa	melissa etheridge	897
00000c289a1829a808ac09c00daf10bc3c4e223b	3d6bbeb7-f90e-4d10-b440-e153c0d10b53	elvenking	717
00000c289a1829a808ac09c00daf10bc3c4e223b	bbd2ffd7-17f4-4506-8572-c1ea58c3f9a8	juliette & the licks	706

We need our inputs to only consist of user and item interactions, so we create a new CSV file which only contains the first column (the hashed userid) and the third column (the artist name) from the original data:

$ cat lastfm-dataset-360K/usersha1-artmbid-artname-plays.tsv|cut -f1,3 > plays.csv

Now the CSV file is in the correct format:

$ head plays.csv 
00000c289a1829a808ac09c00daf10bc3c4e223b	betty blowtorch
00000c289a1829a808ac09c00daf10bc3c4e223b	die Ärzte
00000c289a1829a808ac09c00daf10bc3c4e223b	melissa etheridge
00000c289a1829a808ac09c00daf10bc3c4e223b	elvenking
00000c289a1829a808ac09c00daf10bc3c4e223b	juliette & the licks

Next, we invoke recoreco, point it to the CSV file as input and ask it to write the output to a file called artists.json. It will read the CSV file twice, once for computing some statistics of the data, and a second time for computing the actual item-to-item recommendations. Note that recoreco is pretty fast, the computation takes less than a minute on my machine.

$ recoreco --inputfile=plays.csv --outputfile=artists.json

Reading plays.csv to compute data statistics (pass 1/2)
Found 17535655 interactions between 358868 users and 292365 items.
Reading plays.csv to compute 10 item indicators per item (pass 2/2)
194996130 cooccurrences observed, 34015ms training time, 292365 items rescored
Writing indicators...

The file artists.json now contains the results of the computation. Let's have a look at some artist recommendations using the JSON processor jq.

Who is strongly associated with Michael Jackson?

$ jq 'select(.for_item=="michael jackson")' artists.json

{
  "for_item": "michael jackson",
  "indicated_items": [
    "justin timberlake",
    "queen",
    "kanye west",
    "amy winehouse",
    "britney spears",
    "madonna",
    "rihanna",
    "beyoncé",
    "daft punk",
    "u2"
  ]
}

One of my favorite bands is Hot Water Music, lets see bands that people associate with them:

$ jq 'select(.for_item=="hot water music")' artists.json

{
  "for_item": "hot water music",
  "indicated_items": [
    "lifetime",
    "the get up kids",
    "the lawrence arms",
    "the gaslight anthem",
    "dillinger four",
    "propagandhi",
    "the bouncing souls",
    "strike anywhere",
    "jawbreaker",
    "chuck ragan"
  ]
}

And finally, we look for artists similar to Paco de Lucia in homage to Ted's days of building search engines for Veoh :)

$ jq 'select(.for_item=="paco de lucia")' artists.json

{
  "for_item": "paco de lucia",
  "indicated_items": [
    "miguel poveda",
    "cserhati zsuzsa",
    "ramón veloz",
    "szarka tamás",
    "camaron de la isla",
    "cseh tamás - másik jános",
    "duquende",
    "amr diab",
    "chuck brown & eva cassidy",
    "keympa"
  ]
}

Programmatic Usage

recoreco can also be included as a library in your rust program. We provide a basic example on how to do this. Be sure to checkout the documentation for further details.