Batch-Running Recommenders

The functions in lenskit.batch enable you to generate many recommendations or predictions at the same time, useful for evaluations and experiments.

The batch functions can parallelize over users with the optional n_jobs parameter, or the LK_NUM_PROCS environment variable.


Scripts calling the batch recommendation or prediction facilites must be protected; that is, they should not directly perform their work when run, but should define functions and call a main function when run as a script, with a block like this at the end of the file:

def main():
    # do the actual work

if __name__ == '__main__':

If you are using the batch functions from a Jupyter notbook, you should be fine - the Jupyter programs are appropriately protected.


lenskit.batch.recommend(algo, users, n, candidates=None, *, n_jobs=None, **kwargs)

Batch-recommend for multiple users. The provided algorithm should be a algorithms.Recommender.

  • algo – the algorithm

  • users (array-like) – the users to recommend for

  • n (int) – the number of recommendations to generate (None for unlimited)

  • candidates – the users’ candidate sets. This can be a function, in which case it will be passed each user ID; it can also be a dictionary, in which case user IDs will be looked up in it. Pass None to use the recommender’s built-in candidate selector (usually recommended).

  • n_jobs (int) – The number of processes to use for parallel recommendations. Passed to lenskit.util.parallel.invoker().


A frame with at least the columns user, rank, and item; possibly also score, and any other columns returned by the recommender.

Rating Prediction

lenskit.batch.predict(algo, pairs, *, n_jobs=None, **kwargs)

Generate predictions for user-item pairs. The provided algorithm should be a algorithms.Predictor or a function of two arguments: the user ID and a list of item IDs. It should return a dictionary or a pandas.Series mapping item IDs to predictions.

To use this function, provide a pre-fit algorithm:

>>> from lenskit.algorithms.bias import Bias
>>> from lenskit.metrics.predict import rmse
>>> from lenskit import datasets
>>> ratings = datasets.MovieLens('data/ml-latest-small').ratings
>>> bias = Bias()
<lenskit.algorithms.bias.Bias object at ...>
>>> preds = predict(bias, ratings[-1000:])
>>> preds.head()
       user  item  rating   timestamp  prediction
99004   664  8361     3.0  1393891425    3.288286
99005   664  8528     3.5  1393891047    3.559119
99006   664  8529     4.0  1393891173    3.573008
99007   664  8636     4.0  1393891175    3.846268
99008   664  8641     4.5  1393890852    3.710635
>>> rmse(preds['prediction'], preds['rating'])
  • algo (lenskit.algorithms.Predictor) – A rating predictor function or algorithm.

  • pairs (pandas.DataFrame) – A data frame of (user, item) pairs to predict for. If this frame also contains a rating column, it will be included in the result.

  • n_jobs (int) – The number of processes to use for parallel batch prediction. Passed to lenskit.util.parallel.invoker().


a frame with columns user, item, and prediction containing the prediction results. If pairs contains a rating column, this result will also contain a rating column.

Return type


Isolated Training

This function isn’t a batch function per se, as it doesn’t perform multiple operations, but it is primarily useful with batch operations. The train_isolated() function trains an algorithm in a subprocess, so all temporary resources are released by virtue of the training process exiting. It returns a shared memory serialization of the trained model, which can be passed directly to recommend() or predict() in lieu of an algorithm object, to reduce the total memory consumption.

Example usage:

algo = BiasedMF(50)
algo = Recommender.adapt(algo)
algo = batch.train_isolated(algo, train_ratings)
preds = batch.predict(algo, test_ratings)
lenskit.batch.train_isolated(algo, ratings, *, file=None, **kwargs)

Train an algorithm in a subprocess to isolate the training process. This function spawns a subprocess (in the same way that LensKit’s multiprocessing support does), calls on it, and serializes the result for shared-memory use.

Training the algorithm in a single-purpose subprocess makes sure that any training resources, such as TensorFlow sessions, are cleaned up by virtue of the process terminating when model training is completed. It can also reduce memory use, because the original trained model and the shared memory version are not in memory at the same time. While the batch functions use shared memory to reduce memory overhead for parallel processing, naive use of these functions will still have 2 copies of the model in memory, the shared one and the original, because the sharing process does not tear down the original model. Training in a subprocess solves this problem elegantly.

  • algo (lenskit.algorithms.Algorithm) – The algorithm to train.

  • ratings (pandas.DataFrame) – The rating data.

  • file (str or pathlib.Path or None) – The file in which to save the trained model. If None, uses a default file path or shared memory.

  • kwargs (dict) – Additional named parameters to


The saved model object. This is the owner, so it needs to be closed when finished to free resources.

Return type


Scripting Evaluation

The MultiEval class is useful to build scripts that evaluate multiple algorithms or algorithm variants, simultaneously, across multiple data sets. It can extract parameters from algorithms and include them in the output, useful for hyperparameter search.

For example:

from lenskit.batch import MultiEval
from lenskit.crossfold import partition_users, SampleN
from lenskit.algorithms import basic, als
from lenskit.datasets import MovieLens
from lenskit import topn
import pandas as pd

ml = MovieLens('ml-latest-small')

eval = MultiEval('my-eval', recommend=20)
eval.add_datasets(partition_users(ml.ratings, 5, SampleN(5)), name='ML-Small')
eval.add_algorithms(basic.Popular(), name='Pop')
eval.add_algorithms([als.BiasedMF(f) for f in [20, 30, 40, 50]],
                    attrs=['features'], name='ALS')

The my-eval/runs.csv file will then contain the results of running these algorithms on this data set. A more complete example is available in the MultiEval notebook.

class lenskit.batch.MultiEval(path, *, predict=True, recommend=100, candidates=None, save_models=False, eval_n_jobs=None, combine=True, **kwargs)

Bases: object

A runner for carrying out multiple evaluations, such as parameter sweeps.

  • path (str or pathlib.Path) – the working directory for this evaluation. It will be created if it does not exist.

  • predict (bool) – whether to generate rating predictions.

  • recommend (int) – the number of recommendations to generate per user. Any false-y value (None, False, 0) will disable top-n. The literal value True will generate recommendation lists of unlimited size.

  • candidates (function) – the default candidate set generator for recommendations. It should take the training data and return a candidate generator, itself a function mapping user IDs to candidate sets. Pass None to use the default candidate set configured for each algorithm (recommended).

  • save_models (bool or str) – save individual estimated models to disk. If True, models are pickled to .pkl files; if 'gzip', they are pickled to gzip-compressed .pkl.gz files.

  • eval_n_jobs (int or None) – Value to pass to the n_jobs parameter in lenskit.batch.predict() and lenskit.batch.recommend().

  • combine (bool) – whether to combine output; if False, output will be left in separate files, if True, it will be in a single set of files (runs, recommendations, and predictions).

add_algorithms(algos, attrs=[], **kwargs)

Add one or more algorithms to the run.

  • algos (algorithm or list) – the algorithm(s) to add.

  • attrs (list of str) – a list of attributes to extract from the algorithm objects and include in the run descriptions.

  • kwargs – additional attributes to include in the run descriptions.

add_datasets(data, name=None, candidates=None, **kwargs)

Add one or more datasets to the run.

  • data

    The input data set(s) to run. Can be one of the following:

    • A tuple of (train, test) data.

    • An iterable of (train, test) pairs, in which case the iterable is not consumed until it is needed.

    • A function yielding either of the above, to defer data load until it is needed.

    Data can be either data frames or paths; paths are loaded after detection using util.read_df_detect().

  • kwargs – additional attributes pertaining to these data sets.


Persist the data for an experiment, replacing in-memory data sets with file names. Once this has been called, the sweep can be pickled.


Get the number of runs in this evaluation.

run(runs=None, *, progress=None)

Run the evaluation.

  • runs (int or set-like) – If provided, a specific set of runs to run. Useful for splitting an experiment into individual runs. This is a set of 1-based run IDs, not 0-based indexes.

  • progress – A tqdm.tqdm()-compatible progress function.


Collect the results from non-combined runs into combined output files.