Algorithm Implementation Tips

Implementing algorithms is fun, but there are a few things that are good to keep in mind.

In general, development follows the following:

  1. Correct

  2. Clear

  3. Fast

In that order. Further, we always want LensKit to be usable in an easy fashion. Code implementing algorithms, however, may be quite complex in order to achieve good performance.


We use Numba to optimize critical code paths and provide parallelism in a number of cases, such as ALS training. See the ALS source code for examples.

We also directly use MKL sparse matrix routines when available for some operations. Whenever this is done in the main LensKit code base, however, we also provide fallback implementations when the MKL is not available. The k-NN recommenders both demonstrate different versions of this. The _mkl_ops module exposes MKL operations; we implement them through C wrappers in the mkl_ops.c file, that are then called through FFI. This extra layer is because the raw MKL calls are quite complex to call via FFI, and are not particularly amenable to use with Numba. We re-expose simplified interfaces that are also usable with Numba.

Pickling and Sharing

LensKit uses Python pickling (or JobLib’s modified pickling in joblib.dump()) quite a bit to save and reload models and to share model data between concurrent processes. This generally just works, and you don’t need to implement any particular save/load logic in order to have your algorithm be savable and sharable.

There are a few exceptions, though.

If your algorithm updates state after fitting, this should not be pickled. An example of this would be caching predictions or recommendations to save time in subsequent calls. Only the model parameters and estimated parameters should be pickled. If you have caches or other ephemeral structures, override __getstate__ and __setstate__ to exclude them from the saved data and to initialize caches to empty values on unpickling.

If your model excludes secondary data structures from pickling, such as a reverse index of user-item interactions, then you should only exclude them when pickling for serialization. When pickling for model sharing (see lenskit.sharing.in_share_context()), you should include the derived structures so they can also be shared.

If your algorithm uses subsidiary models as a part of the training process, but does not need them for prediction or recommendation, then consider overriding __getstate__ to remove the underlying model or replace it with a cloned copy (with lenskit.util.clone()) to reduce serialized disk space (and deserialized memory use).

Random Number Generation

LensKit uses seedbank for managing RNG seeds and constructing random number generation.

In general, algorithms using randomization should have an rng_spec parameter that takes a seed or RNG, and pass this to seedbank.numpy_rng() to get a random number generator. Algorithms that use randomness at predict or recommendation time, not just training time, should support the value 'user' for the rng parameter, and if it is passed, derive a new seed for each user using seedbank.derive_seed() to allow reproducibility in the face of parallelism for common experimental designs. lenskit.util.derivable_rng() automates this logic.

Memory Map Friendliness

LensKit uses joblib.Parallel to parallelize internal operations (when it isn’t using Numba). Joblib is pretty good about using shared memory to minimize memory overhead in parallel computations, and LensKit has some tricks to maximize this use. However, it does require a bit of attention in your algorithm implementation.

The easiest way to make this fail is to use many small NumPy or Pandas data structures. If you have a dictionary of np.ndarray objects, for instance, it will cause a problem. This is because each array will be memory-mapped, and each map will reopen the file. Having too many active open files will cause your process to run out of file descriptors on many systems. Keep your object count to a small, ideally fixed number; in lenskit.algorithms.basic.UnratedItemSelector, we do this by storing user and item indexes along with a matrix.CSR containing the items rated by each user. The old implementation had a dictionary mapping user IDs to ``ndarray``s with each user’s rated items. This is a change from \(|U|+1\) arrays to 5 arrays.