license: apache-2.0
Dataset Card for Click Reco Slates
Data structure. The data contain three files that contain click logs. These are train.parquet, val.parquet (validate) and special.parquet.
The file format is:
R_bar_r contains 0 for no click, or a i for a click on item i.
The interest contains the interests of the user.
The engagement contains other features that may help predict if the user clicks, one_hot_engagement is simply a one hot encoded representation of the engagements.
In addition the following embedding files are included:
products_embeddings.pt interest_embeddings.pt
which contain embeddings that can be indexed by the entries in slates (products_embbeddings.pt) or by the keyword (interest_embeddings.pt).
The train.parquet is intended for training and the val.parquet for validation - they are a randomized split. The special set special.parquet contains slates that are on a special list that are never seen in train and validate.
The motivation of this is that a click model for reco needs not to explain the data well as with likelihood, but to be capable of selecting new actions.