lenskit | LensKit recommender toolkit

LensKit requires the data set to contain your users' data, unless you use the item-based recommenders/scorers. However, model training and recommendation/prediction can use different data sets - LensKit just assumes that you've stored user preferences in LensKIt's database before generating predictions.

Some algorithms (e.g. FunkSVD) ignore user data that isn't in the model. Others (item-item and user-user) make use of current user data in the data access object.

In LensKit 3, the way to work with custom data is to directly use Entity objects. Users, items, and ratings, are just entities; the User, Item, and Rating classes are view classes that provide convenience access to common attributes, but the base Entity interface allows you to use arbitrary attributes. The data model is documented at https://lenskit.gitbooks.io/lenskit-manual/basics/data-model.html; it is also possible to create your own view classes, but this is not currently documented.

To use JDBC, you will need to reimplement the DataAccessObject interface on top of JDBC. Building a reference implementation of this functionality is on the TODO list, but has not yet been done.

This goodItems should work:

user.testHistory.findAll({ it instanceof org.lenskit.data.ratings.Rating && it.value >= 4 })*.itemId.toSet()

There are two pieces:'

1. Implement your new metric by extendingTopNMetric (or ListOnlyTopNMetric). The X type parameter is the type of context object that is used to keep track of a particular evaluation experiment (algorithm + data set combo). You will probably need to extract whatever data you're using to compute ILS from the recommender in createContext and save it in your context object. The context class is usually a static inner class of the metric class.

2. Create a properties file, META-INF/lenskit/topn-metrics.properties, that will show in your classpath (put it under src/main/resources in a standard Maven or Gradle project layout) to associate a name with your metric's class. There you write

   ils=my.package.ILSTOpNMetric

This can all live in the same general project as your algorithm code.