FAQ

1. What kind of features should I use: Dense or Sparse?

   It depends. In most cases you’re better off using SparseFeatures because they are better suited to categorical features. Even if you have a feature with real numerical values you’re often better off if you binarize or discretize it. But there are exceptions to this rule, e.g. ALS features.

2. How do I calculate several metrics at once?

   Use function calc_metrics. It allows to calculate a batch of metrics more efficiently. It’s similar to reports from sklearn.

3. What is the benefit of model wrappers?

   They all have the same set of parameters allowing for easier usage. They also provide extension of existing functionality, such as allowing to filters to eliminate items that has already been seen, whitelist, features in ALS, I2I. Wrappers have unified interface of output that is easy to use as input to calculate metrics. They also allowed to speed up performance of some models.

4. What is the benefit of using Dataset?

   It’s an easy-to-use wrapping of interactions, features and mapping between item and user ids in feature sets and those in interaction matrix.

5. Why do I need to pass Dataset object as an argument to method recommend?

   It conceals mapping between internal and external user and item ids. Additionally it allows to filter out items that users have already seen. Some models, such as LightFM or DSSM, require to pass features.

6. Should the same Dataset object be used for fitting of a model and for inference of recommendations?

   It almost always has to be exactly the same Dataset object.

   One of possible exceptions is if during the fitting stage you use both viewing and purchase of an item as a positive event but you want exempt an item from being recommended only if it was purchased. In this case you should pass all interactions to train a model and only purchases to infer recommendations.

   Another exception is if a model requires to pass features to infer recommendations and values of those features have changed.