LensKit-Auto

LensKit-Auto is built as a wrapper around the Python LensKit recommender-system library. It automates algorithm selection and hyper parameter optimization an can build ensemble models based on the LensKit models.

Resources

• Documentation: LensKit-Auto Documentation
• RecSys23 Demo Video: RecSys23 Demo Video
• RecSys Demo: RecSys Demo

Installation

LensKit-Auto requires Python 3.12 or newer.
Install it and all dependencies with:

pip install lenskit-auto

Recommended Environment Setup

On Linux:

python3 -m venv lenskit-auto-env
source lenskit-auto-env/bin/activate
pip install lenskit-auto

On Windows:

py -3.12 -m venv lenskit-auto-env
lenskit-auto-env\Scripts\activate
pip install lenskit-auto

Tip: You can replace lenskit-auto-env with any environment name you prefer.

Installing the DeepCave package

DeepCave is a tool to visualize the runs LensKit-Auto takes. LensKit-Auto uses a newer Pyhon version than DeepCave, however, in our experience DeepCave still works with the newer version. But due to the version differences, we need to install DeepCave manually. For this we use these two commands:

pip install deepcave --ignore-requires-python
pip install matplotlib=3.10.3
pip install plotly==6.1.1

The first command installs the DeepCave package and ignores its Python version requirements. During the DeepCave installation, our installed matplotlib package version is downgraded, so we manually need to reinstall the correct version again with the second command. DeepCave also installs an older version version of plotly, that we need to upgrade too.

Getting Started

LensKit-Auto is built as a wrapper around the Python LensKit recommender-system library. It automates algorithm selection and hyper parameter optimization and can build ensemble models based on LensKit models.

Standard use-case

In the standard use-case you just need to call a single function to get the best performing model for your dataset. It is either

from lkauto.lkauto import get_best_recommender_model

get_best_recommender_model(train=train_split)

for the recommendation use-case or

from lkauto.lkauto import get_best_prediction_model

get_best_prediction_model(train=train_split)

for the prediction use-case

Examples and Advanced Use-Cases

LensKit-Auto allows three application scenarios:

Note: All application scenarios apply to Top-N ranking prediction and rating prediction use-cases.

• Scenario 1: LensKit-Auto performs combined algorithm selection and hyperparameter optimization for a given dataset.
• Scenario 2: LensKit-Auto performs hyperparameter optimization on a single algorithm for a given dataset.
• Scenario 3: LensKit-Auto performs combined algorithm selection and hyperparameter optimization for a specified set of algorithms and/or different hyperparameter ranges for the provided dataset.

In order to take advantage of LensKit-Auto, a developer needs to read in a dataset. The load_movielens() function can be used to load a MovieLens dataset for example.

from lenskit.data import load_movielens

ml100k = load_movielens('path_to_file')

Furthermore, it is suggested, that we take advantage of the Filer to control the LensKit-Auto output

from lkauto.utils.filer import Filer

filer = Filer('output/')

Top-N ranking prediction

First, we need to split the data in a train and test split to evaluate our model. The train-test splits can be performed based on data rows or user data. For the rating prediction example we are splitting the data based on user data.

from lenskit.splitting import crossfold_users, SampleN
from lenskit.batch import recommend
from lenskit.metrics import RunAnalysis, NDCG
from lkauto.lkauto import get_best_recommender_model

# User based data-split
for split in crossfold_users(ml100k, 2, SampleN(5)):
    train_split = split.train
    test_split = split.test

    # INSERT SCENARIO CODE HERE
    # See Scenario 1, 2, and 3 sections below

    # recommend
    recs = recommend(model, test_split.keys())

    # initialize analysis
    analysis = RunAnalysis()
    # add ndcg metric
    analysis.add_metric(NDCG())
    # evaluate recommendations against the test interactions
    scores = analysis.measure(recs, test_split)

Rating Prediction

First, we need to split the data in a train and test split to evaluate our model. The train-test splits can be performed based on data rows or user data. For the rating prediction example we are splitting the data based on the data rows. The Top-N ranking predicion example showcases the data-split based on user data.

from lenskit.splitting import sample_records
from lenskit.metrics import RunAnalysis, RMSE
from lkauto.lkauto import get_best_prediction_model
from lkauto.utils.pred_and_rec_functions import predict

# random split: hold out 25% of interactions for testing
test_size = int(ml100k.interaction_count * 0.25)
split = sample_records(ml100k, size=test_size)
train_split = split.train
test_split = split.test

# INSERT SCENARIO CODE HERE
# See Scenario 1, 2, and 3 sections below

# generate rating predictions for the held-out interactions
# here we need to use the wrapper predict() and not the lenskit.batch.predict()
predictions = predict(model, test_split)

# initialize analysis
analysis = RunAnalysis()
# add rmse metric
analysis.add_metric(RMSE())
# evaluate recommendations against the test interactions
scores = analysis.measure(predictions, test_split)

Application Scenarios

The following scenarios can be used in both Top-N ranking prediction and rating prediction use cases. Simply insert the scenario code where indicated by # INSERT SCENARIO CODE HERE in the examples above.

Scenario 1: Fully Automated Model Selection & Hyperparameter Optimization

Scenario 1 describes the fully automated combined algorithm selection and hyperparameter optimization (CASH problem). This scenario is recommended for inexperienced developers who have no or little experience in model selection.

LensKit-Auto performs the combined algorithm selection and hyperparameter optimization with a single function call.

model, config = get_best_recommender_model(train=train_split, filer=filer)

Note: As described above, the get_best_recommender_model() is used for Top-N ranking prediction. If you want to find a predictor instead of a recommender, replace the function call with get_best_prediction_model()

The get_best_recommender_model() or get_best_prediction_model() function call will return the best performing model, with tuned hyperparameters and a configuration dictionary that contains the best configuration found during the optimization. In the Scenario 1 use-case the model is chosen out of all LensKit algorithms with hyperparameters within the LensKit-Auto default hyperparameter range.

We can use the model in the exact same way like a regular LensKit model. For rating prediction with ensemble models, use the predict() wrapper function from lkauto.utils.pred_and_rec_functions instead of lenskit.batch.predict() directly (as shown in the Rating Prediction example above).

Setting the save parameter to True enables lenskit-auto to save the trained model and configuration to the ouput directory specified by the filer. The default value of save is True, so that we only have to set it to False if we do not want to save the model and configuration.

Scenario 2: Single-Algorithm Hyperparameter Optimization

In Scenario 2 we are going to perform hyperparameter optimization on a single algorithm. First we need to define our custom configuration space with just a single algorithm included.

from ConfigSpace import Constant
from lkauto.algorithms.item_knn import ItemItem

# initialize ItemItem ConfigurationSpace
cs = ItemItem.get_default_configspace()
# add algorithm name as a constant (in this case ItemItem algorithm)
cs.add([Constant("algo", "ItemItem")])
# set a random seed for reproducible results
cs.seed(42)

# Provide the ItemItem ConfigurationSpace to the get_best_recommender_model function.
model, config = get_best_recommender_model(train=train_split, filer=filer, cs=cs)

Note: As described above, the get_best_recommender_model() is used for Top-N ranking prediction. If you want to find a predictor instead of a recommender, replace the function call with get_best_prediction_model()

The get_best_recommender_model() or get_best_prediction_model() function call will return the best performing ItemItem model. Besides the model, the get_best_recommender_model() function returns a configuration dictionary with the best configuration found during the optimization.

Scenario 3: Custom Search Space Model Selection & Hyperparameter Optimization

Scenario 3 describes the automated combined algorithm selection and hyperparameter optimization of a custom configuration space. A developer that wants to use Scenario 3 needs to provide hyperparameter ranges for the hyperparameter optimization process.

First, a parent-ConfigurationSpace needs to be initialized. All algorithm names need to be added to the parent-ConfigurationSpace categorical algo hyperparameter.

from ConfigSpace import ConfigurationSpace, Categorical

# initialize parent ConfigurationSpace
parent_cs = ConfigurationSpace()
# set a random seed for reproducible results
parent_cs.seed(42)
# add algorithm names as a constant
parent_cs.add([Categorical("algo", ["ItemItem", "UserUser"])])

Afterward, we need to build the ItemItem and UserUser sub-ConfigurationSpace.

We can use the default sub-ConfigurationSpace from LensKit-Auto and add it to the parent-ConfigurationSpace:

from lkauto.algorithms.item_knn import ItemItem

# get default ItemItem ConfigurationSpace
item_item_cs = ItemItem.get_default_configspace()

# Add sub-ConfigurationSpace to parent-ConfigurationSpace
parent_cs.add_configuration_space(
    prefix="ItemItem",
    delimiter=":",
    configuration_space=item_item_cs,
    parent_hyperparameter={"parent": parent_cs["algo"], "value": "ItemItem"},
)

Or we can build our own ConfigurationSpace for a specific algorithm.

from ConfigSpace import ConfigurationSpace
from ConfigSpace import Integer, Float, Constant

# first we initialize hyperparameter objects for all hyperparameters that we want to optimize
nnbrs = Constant('nnbrs', 1000)
min_nbrs = Integer('min_nbrs', bounds=(1, 50), default=1)
min_sim = Float('min_sim', bounds=(0, 0.1), default=0)

# Then, we initialize the sub-ConfigurationSpace and add the hyperparameters to it
user_user_cs = ConfigurationSpace()
user_user_cs.add([nnbrs, min_nbrs, min_sim])

# Last, we add the user_user_cs to the parent-ConfigurationSpace 

parent_cs.add_configuration_space(
    prefix="UserUser",
    delimiter=":",
    configuration_space=user_user_cs,
    parent_hyperparameter={"parent": parent_cs["algo"], "value": "UserUser"},
)

After creating the parent-ConfigurationSpace, we can use it in the same way like Scenario 2

# Provide the parent-ConfigurationSpace to the get_best_recommender_model function. 
model, config = get_best_recommender_model(train=train_split, filer=filer, cs=parent_cs)

Note: As described above, the get_best_recommender_model() is used for Top-N ranking prediction. If you want to find a predictor instead of a recommender, replace the function call with get_best_prediction_model()

Experiments on Default Configurations

The experiments to gather some hyperparameters for LensKit-Autos default configuration are described here: Experiments