Illoominate is a library that implements the KMC-Shapley algorithm for computing data importance scores in recommender systems. The KMC-Shapley algorithm leverages the sparsity and nearest-neighbor structure of sequential kNN-based recommendation models to efficiently compute Data Shapley values (DSV) and leave-one-out (LOO) scores. This algorithmic approach enables scalable data debugging and importance analysis for real-world datasets with millions of interactions in session-based and next-basket recommendation tasks.
This repository contains the official code for the Illoominate framework, which accompanies our paper accepted at RecSys 2025. The paper is available with open access at https://dl.acm.org/doi/10.1145/3705328.3748049.
Citation:
@inproceedings{kersbergen-2025-kmc-shapley,
author = {Kersbergen, Barrie and Sprangers, Olivier and Karla{\v s}, Bojan and de Rijke, Maarten and Schelter, Sebastian},
booktitle = {RecSys 2025: 19th ACM Conference on Recommender Systems},
date-added = {2025-07-03 20:57:23 +0200},
date-modified = {2025-07-03 21:01:57 +0200},
month = {September},
publisher = {ACM},
title = {Scalable Data Debugging for Neighborhood-based Recommendation with Data Shapley Values},
year = {2025}}- KMC-Shapley Algorithm: Implements the novel K-Monte Carlo Shapley algorithm that exploits the sparsity and nearest-neighbor structure of kNN-based models to efficiently compute Data Shapley values.
- Scalable Data Debugging: Enables data importance analysis on large datasets with millions of interactions through algorithmic optimization rather than just computational efficiency.
- Multiple Model Support: Works with sequential kNN-based recommendation models including VMIS-kNN (session-based) and TIFU-kNN (next-basket), supporting popular metrics such as MRR, NDCG, Recall, F1 etc.
- Practical Applications: Designed for real-world use cases including data debugging, quality assessment, data pruning, and sustainable recommendation systems.
Illoominate implements the KMC-Shapley algorithm, which is specifically designed for sequential kNN-based recommendation models. The algorithm's key insight is that most data points in kNN-based systems only influence a small subset of predictions through their nearest-neighbor relationships. By exploiting this sparsity, KMC-Shapley avoids redundant computations and enables scalable Data Shapley value estimation.
The library provides a Python frontend with Rust backend implementation, supporting KNN-based models VMIS-kNN (session-based) and TIFU-kNN (next-basket) for real-world recommendation scenarios.
By leveraging the KMC-Shapley algorithm, Illoominate enables data scientists and engineers to:
- Debug potentially corrupted data through efficient importance scoring
- Improve recommendation quality by identifying the most impactful data points
- Prune training data for sustainable and efficient recommendation systems
Illoominate is available via PyPI.
pip install illoominateEnsure Python >= 3.10 is installed. We provide precompiled binaries for Linux, Windows and macOS.
It is recommended to install and run Illoominate from a virtual environment. If you are using a virtual environment, activate it before running the installation command.
python -m venv venv # Create the virtual environment (Linux/macOS/Windows)
source venv/bin/activate # Activate the virtualenv (Linux/macOS)
venv\Scripts\activate # Activate the virtualenv (Windows)
# Load training and validation datasets
train_df = pd.read_csv('data/tafeng/processed/train.csv', sep='\t')
validation_df = pd.read_csv('data/tafeng/processed/valid.csv', sep='\t')
# Data Leave-One-Out values
loo_values = illoominate.data_loo_values(
train_df=train_df,
validation_df=validation_df,
model='tifu',
metric='ndcg@10',
params={'m':7, 'k':100, 'r_b': 0.9, 'r_g': 0.7, 'alpha': 0.7, 'seed': 42},
)
# Visualize the distribution of Data Leave-One-Out Values
plt.hist(shapley_values['score'], density=False, bins=100)
plt.title('Distribution of Data LOO Values')
plt.yscale('log')
plt.ylabel('Frequency')
plt.xlabel('Data Leave-One-Out Values')
plt.savefig('images/loo.png', dpi=300)
plt.show()
Illoominate computes Data Shapley values to assess the contribution of each data point to the recommendation performance. Below is an example using the public Now Playing 1M dataset.
import illoominate
import matplotlib.pyplot as plt
import pandas as pd
# Load training and validation datasets
train_df = pd.read_csv("data/nowplaying1m/train.csv", sep='\t')
validation_df = pd.read_csv("data/nowplaying1m/valid.csv", sep='\t')
# Compute Data Shapley values
shapley_values = illoominate.data_shapley_values(
train_df=train_df,
validation_df=validation_df,
model='vmis', # Model to be used (e.g., 'vmis' for VMIS-kNN)
metric='mrr@20', # Evaluation metric (e.g., Mean Reciprocal Rank at 20)
params={'m':100, 'k':100, 'seed': 42}, # Model-specific parameters
)
# Visualize the distribution of Data Shapley values
plt.hist(shapley_values['score'], density=False, bins=100)
plt.title('Distribution of Data Shapley Values')
plt.yscale('log')
plt.ylabel('Frequency')
plt.xlabel('Data Shapley Values')
plt.savefig('images/shapley.png', dpi=300)
plt.show()
# Identify potentially corrupted sessions
negative = shapley_values[shapley_values.score < 0]
corrupt_sessions = train_df.merge(negative, on='session_id')
The distribution of Data Shapley values can be visualized or used for further analysis.
print(corrupt_sessions)
session_id item_id timestamp score
0 5076 64 1585507853 -2.931978e-05
1 13946 119 1584189394 -2.606203e-05
2 13951 173 1585417176 -6.515507e-06
3 3090 199 1584196605 -2.393995e-05
4 5076 205 1585507872 -2.931978e-05
... ... ... ... ...
956 13951 5860 1585416925 -6.515507e-06
957 447 3786 1584448579 -5.092383e-06
958 7573 14467 1584450303 -7.107826e-07
959 5123 47 1584808576 -4.295939e-07
960 11339 4855 1585391332 -1.579517e-06
961 rows × 4 columnsTo compute Data Shapley values for next-basket recommendations, use the Tafeng dataset.
# Load training and validation datasets
train_df = pd.read_csv('data/tafeng/processed/train.csv', sep='\t')
validation_df = pd.read_csv('data/tafeng/processed/valid.csv', sep='\t')
# Compute Data Shapley values
shapley_values = illoominate.data_shapley_values(
train_df=train_df,
validation_df=validation_df,
model='tifu',
metric='ndcg@20',
params={'m': 7, 'k': 100, 'r_b': 0.9, 'r_g': 0.7, 'alpha': 0.7, 'seed': 42, 'convergence_threshold': .1},
)
# Visualize the distribution of Data Shapley values
plt.hist(shapley_values['score'], density=False, bins=100)
plt.title('Distribution of Data Shapley Values')
plt.yscale('log')
plt.ylabel('Frequency')
plt.xlabel('Data Shapley Values')
plt.savefig('images/shapley.png', dpi=300)
plt.show()
Illoominate supports metrics to include a sustainability term that expresses the number of sustainable products in a given recommendation. SustainableMRR@t as 0.8·MRR@t + 0.2· st . This utility combines the MRR@t with the “sustainability coverage term” st , where s denotes the number of sustainable items among the t recommended items.
The function call remains the same, you only change the metric to SustainableMRR, SustainableNDCG or st (sustainability coverage term) and provide a list of items that are considered sustainable.
import illoominate
import matplotlib.pyplot as plt
import pandas as pd
train_df = pd.read_csv('data/rsc15_100k/processed/train.csv', sep='\t')
validation_df = pd.read_csv('data/rsc15_100k/processed/valid.csv', sep='\t')
# rsc15 items considered sustainable. (Randomly chosen for this dataset)
sustainable_df = pd.read_csv('data/rsc15_100k/processed/sustainable.csv', sep='\t')
importance = illoominate.data_loo_values(
train_df=train_df,
validation_df=validation_df,
model='vmis',
metric='sustainablemrr@20',
params={'m':500, 'k':100, 'seed': 42},
sustainable_df=sustainable_df,
)
plt.hist(importance['score'], density=False, bins=100)
plt.title('Distribution of Data Leave-One-Out Values')
plt.yscale('log')
plt.ylabel('Frequency')
plt.xlabel('Data Leave-One-Out Values')
plt.savefig('data/rsc15_100k/processed/loo_responsiblemrr.png', dpi=300)
plt.show()
# Prune the training data
threshold = importance['score'].quantile(0.05) # 5th percentile threshold
filtered_importance_values = importance[importance['score'] >= threshold]
train_df_pruned = train_df.merge(filtered_importance_values, on='session_id')
This demonstrates the pruned training dataset, where less impactful or irrelevant interactions have been removed to focus on high-quality data points for model training.
print(train_df_pruned)
session_id item_id timestamp score
0 3 214716935 1.396437e+09 0.000000
1 3 214832672 1.396438e+09 0.000000
2 7 214826835 1.396414e+09 -0.000003
3 7 214826715 1.396414e+09 -0.000003
4 11 214821275 1.396515e+09 0.000040
... ... ... ... ...
47933 31808 214820441 1.396508e+09 0.000000
47934 31812 214662819 1.396365e+09 -0.000002
47935 31812 214836765 1.396365e+09 -0.000002
47936 31812 214836073 1.396365e+09 -0.000002
47937 31812 214662819 1.396365e+09 -0.000002Illoominate allows you to train a kNN-based model and evaluate it directly using Python.
import illoominate
import pandas as pd
# Load training and validation datasets
train_df = pd.read_csv("data/nowplaying1m/train.csv", sep="\t")
validation_df = pd.read_csv("data/nowplaying1m/valid.csv", sep="\t")
# Define model and evaluation parameters
model = 'vmis'
metric = 'mrr@20'
params = {
'm': 500, # session memory
'k': 100, # number of neighbors
'seed': 42 # random seed
}
# Run training and evaluation
scores = illoominate.train_and_evaluate_for_sbr(
train_df=train_df,
validation_df=validation_df,
model=model,
metric=metric,
params=params
)
print(f"Evaluation score ({metric}):", scores['score'][0])You can also evaluate against a separate test set if needed:
validation_scores = illoominate.train_and_evaluate_for_sbr(
train_df=train_df,
validation_df=val_df,
model='vmis',
metric='mrr@20',
params=params
)
test_scores = illoominate.train_and_evaluate_for_sbr(
train_df=train_df,
validation_df=test_df,
model='vmis',
metric='mrr@20',
params=params
)For large session-based datasets, you can avoid loading pandas DataFrames entirely and evaluate directly from delimited files:
import illoominate
scores = illoominate.train_and_evaluate_for_sbr_files(
train_path="data/bol35m/bolcom-clicks-50m_train.txt",
validation_path="data/bol35m/bolcom-clicks-50m_test.txt",
metric="mrr@20",
params={"m": 100, "k": 100, "seed": 42},
sep="\t",
)
print(scores)This approach is designed for large-scale datasets where pandas introduces substantial overhead. In practice, loading data into pandas can require up to ~3× the size of the raw file on disk due to in-memory representations and indexing. By reading directly from delimited files, you avoid this overhead entirely, resulting in significantly lower memory usage.
model (str): Name of the model to use. Supported values:
metric (str): Evaluation metric to calculate importance. Supported values:
mrr@20Mean Reciprocal Rankndcg@20Normalized Discounted Cumulative Gainst@20Sustainability coveragehitrate@20HitRatef1@20F1precision@20Precisionrecall@20Recallsustainablemrr@20Combines the MRR with a Sustainability coverage termsustainablendcg@20Combines the NDCG with a Sustainability coverage term
params (dict): Model specific parameters
sustainable_df (pd.DataFrame):
- This argument is only mandatory for the sustainable related metrics
st,sustainablemrrorsustainablendcg
KMC-Shapley (K Monte Carlo Shapley) is a novel algorithm specifically designed for computing Data Shapley values in kNN-based recommendation systems. The algorithm's core innovation lies in recognizing that most data points in kNN-based models only influence predictions through their nearest-neighbor relationships, creating inherent sparsity in the importance computation.
Key Algorithmic Insights:
- Sparsity Exploitation: The algorithm identifies that most data points only affect a small subset of predictions, avoiding unnecessary computations
- Nearest-Neighbor Structure: Leverages the kNN model's structure to focus importance calculations on relevant neighbor relationships
- Monte Carlo Optimization: Uses strategic sampling to estimate Shapley values while maintaining theoretical guarantees
This algorithmic approach enables scalable Data Shapley value computation on datasets with millions of interactions, making it practical for real-world recommendation systems.
To get started with developing Illoominate or conducting the experiments from the paper, follow these steps:
Requirements:
- Rust >= 1.82
- Python >= 3.10
- Clone the repository:
git clone https://github.com/bkersbergen/illoominate.git
cd illoominate- Create the python wheel by:
pip install -r requirements.txt
maturin develop --releaseThe package runtime dependencies are declared in pyproject.toml, so pip install illoominate, pip install ., and maturin develop all install the Python libraries required at runtime. The requirements.txt file is only for development tooling such as notebooks and local release builds.
The experiments from the paper are implemented in Rust for performance benchmarking. Rust's memory safety and performance characteristics make it well-suited for the computational benchmarks, while the core KMC-Shapley algorithm provides the algorithmic efficiency for scalable Data Shapley value computation.
Prepare a config file for a dataset, describing the model, model parameters and the evaluation metric.
$ cat config.toml
[model]
name = "vmis"
[hpo]
k = 50
m = 500
[metric]
name="MRR"
length=20The software expects the config file for the experiment in the same directory as the data files.
DATA_LOCATION=data/tafeng/processed CONFIG_FILENAME=config.toml cargo run --release --bin removal_impactThis project is licensed under the Apache License 2.0 - see the LICENSE file for details.
© 2026 All rights reserved.
For any queries or further support, please refer to our RecSys 2025 paper. Contributions and discussions are welcome!
Increment the version number in pyproject.toml
Trigger a build using the CI pipeline in Github, via either:
- A push is made to the main branch with a tag matching -rc (e.g., v1.0.0-rc1).
- A pull request is made to the main branch.
- A push occurs on a branch that starts with branch-*.
Download the wheels mentioned in the CI job output and place them in a directory. Navigate to that directory and then
twine upload dist/* -u __token__ -p pypi-SomeSecretAPIToken123This will upload all files in the dist/ directory to PyPI. dist/ is the directory where the wheel files will be located after you unpack the artifact from GitHub Actions.