bind RatingVectorPDAO to EntityCountRatingVectorPDAO
set InteractionEntityType to EntityTypes.forName("purchase")Implicit Feedback
Starting from version 3.0, LensKit provides support for handling implicit feedback, particularly unary or count data such as clicks, purchases, and play counts.
Proxy DAOs
Since most LensKit components think in terms of ratings, the common way to make LensKit use non-rating data is to transform it into rating-like data. Per-user ‘rating vectors’ do not need to be computed from actual explicit ratings; they can be computed or inferred by many means.
LensKit abstracts this through the concept of a proxy DAO: a data access object that transforms data. The RatingVectorPDAO interface defines rating vector proxy DAOs; other LensKit components can use the proxy DAO to look up or stream user rating vectors.
The default implementation retrieves a user’s rating entities and converts them into a rating vector. However, there are two additional built-in options, and you can also create your own:
- EntityCountRatingVectorPDAO
-
This PDAO retrieves entities of a particular type associated with the user, and counts how many appear for each item. This is useful for things such as purchases.
- CountSumRatingVectorPDAO
-
This PDAO retrieves entities of a particular type associated with the user, and sums up the values of the
countinteger attribute on the entities associated with each item. This is useful for batched data such as song play counts.
Both of these PDAOs do require that the entities to be counted or summed contain user and item long-valued attributes storing the user and item IDs, respectively.
Configuring the PDAO
If you want to use purchase entities, each of which has user and item IDs, you can add the following to your algorithm configuration:
With this configuration, any LensKit component that uses a RatingVectorPDAO to look up user rating vectors will use the purchase-based vectors instead of rating-based vectors.
Configuring Algorithms
Some algorithms need to be reconfigured in order to make use of implicit feedback data, as their default settings are not useful for those settings. Common tweaks include:
-
Remove `BiasModel`s — the default rating-based bias computations are not effective for most unary settings.
-
Reconfigure neighborhood scorers — weighted averages do not work with 0–1 data, so neighborhood models need to be reconfigured. For item-item, for example:
bind NeighborhoodScorer to SimilaritySumNeighborhoodScorerUser-user requires a similar change:
bind UserNeighborhoodScorer to SimilaritySumUserNeighborhoodScorer -
Reconfigure normalizers — mean normalization is often ineffective, so don’t use it. Use unit vector normalization or identity normalization instead.