The global scorer for the global recommendation backed by a item-item model.
Algorithm for scoring items given an item-item model and neighborhood scorer.
Compute the similarity between two items.
Compute scores from neighborhoods and score vectors.
Default item scoring algorithm.
Score items based on the basket of items using an item-item CF model.
Score items using an item-item CF model.
Neighborhood scorer that computes the sum of neighborhood similarities.
Neighborhood scorer that computes the weighted average of neighbor scores.
Qualifier for threshold applied to item similarities.
Number of neighbors to retain in the similarity matrix.
The item-item CF implementation is built up of several pieces. The model builder takes the rating data and several parameters and components, such as the similarity function and model size, and computes the similarity matrix. The scorer uses this model to score items.
The basic idea of item-item CF is to compute similarities between items, typically
based on the users that have rated them, and the recommend items similar to the items
that a user likes. The model is then truncated — only the
ModelSize most similar
items are retained for each item – to save space. Neighborhoods are further truncated
when doing recommendation; only the
NeighborhoodSize most similar items that
a user has rated are used to score any given item.
ModelSize is typically
NeighborhoodSize to improve the ability of the recommender to find
When the similarity function is asymmetric (\(s(i,j)=s(j,i)\) does not hold), some care
is needed to make sure that the function is used in the correct direction. Following
Deshpande and Karypis, we use the similarity function as \(s(j,i)\), where \(j\) is the
item the user has purchased or rated and \(i\) is the item that is going to be scored. This
function is then stored in row \(i\) and column \(j\) of the matrix. Rows are then truncated
(so we retain the
ModelSize most similar items for each \(i\)); this direction differs
from Deshpande & Karypis, as row truncation is more efficient & simpler to write within
LensKit's item-item algorithm structure, and performs better in offline tests against the
MovieLens 1M data set
Computation against a particular item the user has rated is done down that item's column.
The scorers and recommenders actually operate on a generic
ItemItemModel, so the
item-based scoring algorithm can be used against other sources of similarity, such as
similarities stored in a database or text index.