We use the fastutil library for optimized primitive collections, such as lists of long or double values and maps of the same. One of the most common types is Long2DoubleMap, which we use to represent sparse vectors mapping user or item IDs to values.

The LensKit utility class Vectors provides utilities for performing vector computations on vectors represented by Long2DoubleMap, such as dot products and Euclidean norms.

We also provide highly memory-efficient, immutable implementations of LongSortedSet and Long2DoubleSortedMap accessible through the LongUtils utility class (see the frozenSet and frozenMap methods). These implementations use sorted arrays to provide \(O(\mathrm{lg}\ n)\) lookup and Vectors methods take advantage of their layout to provide linear-time pairwise operations such as dot products. They also do not have excess overhead due to unused space, unlike the hashtable implementations of maps and sets. Further, when none of the long values exceed the storage space of an int, these implementations will use int arrays instead of long arrays to further reduce memory consumption.

It is a common pattern to use a Long2DoubleOpenHashMap to accumulate results and then convert them to an optimized map using LongUtils.frozenMap for longer-term storage.

We use Google Guava library extensively in LensKit. A number of LensKit classes make use of its Predicate and Function interfaces, and we use its collection and I/O utilities quite a bit.

We use the Grapht dependency injection container to configure and instantiate LensKit recommender components. If you have used Google Guice, many aspects of Grapht should be familiar.

We use Apache Commons Math for linear algebra, including dense vectors and matrices.

There are several other libraries we use in LensKit, including: