![[2602.19339] SplitLight: An Exploratory Toolkit for Recommender Systems Datasets and Splits](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.19339] SplitLight: An Exploratory Toolkit for Recommender Systems Datasets and Splits
Summary
SplitLight is an open-source toolkit designed to enhance the evaluation of recommender systems by providing measurable and comparable data preprocessing and splitting strategies.
Why It Matters
This toolkit addresses critical issues in recommender systems research, such as reproducibility and comparability of results, by allowing researchers to document and analyze their data preparation choices transparently. It supports better decision-making in model evaluation, which is essential for advancing the field.
Key Takeaways
- SplitLight enables measurable and comparable data preprocessing for recommender systems.
- It helps identify issues like temporal leakage and distribution shifts in datasets.
- The toolkit offers both a Python interface and a no-code option for broader accessibility.
- Audit summaries produced by SplitLight enhance transparency in experimental protocols.
- Side-by-side comparisons of splitting strategies improve the reliability of model evaluations.
Computer Science > Information Retrieval arXiv:2602.19339 (cs) [Submitted on 22 Feb 2026] Title:SplitLight: An Exploratory Toolkit for Recommender Systems Datasets and Splits Authors:Anna Volodkevich, Dmitry Anikin, Danil Gusak, Anton Klenitskiy, Evgeny Frolov, Alexey Vasilev View a PDF of the paper titled SplitLight: An Exploratory Toolkit for Recommender Systems Datasets and Splits, by Anna Volodkevich and 4 other authors View PDF HTML (experimental) Abstract:Offline evaluation of recommender systems is often affected by hidden, under-documented choices in data preparation. Seemingly minor decisions in filtering, handling repeats, cold-start treatment, and splitting strategy design can substantially reorder model rankings and undermine reproducibility and cross-paper comparability. In this paper, we introduce SplitLight, an open-source exploratory toolkit that enables researchers and practitioners designing preprocessing and splitting pipelines or reviewing external artifacts to make these decisions measurable, comparable, and reportable. Given an interaction log and derived split subsets, SplitLight analyzes core and temporal dataset statistics, characterizes repeat consumption patterns and timestamp anomalies, and diagnoses split validity, including temporal leakage, cold-user/item exposure, and distribution shifts. SplitLight further allows side-by-side comparison of alternative splitting strategies through comprehensive aggregated summaries and interactive visualizat...
