BNRS Replication

Replication repository for:

Kiddle, R., Welbers, K., Kroon, A., & Trilling, D. (2026). Bisociative Pivoting as a Method for Moderating User Response to Text Similarity in Next-item News Recommendation. ACM Transactions on Recommender Systems.

Layout

• bnrs_algorithm/ — code to produce the recommendation types used in the paper (with a demo dataset: rjac/all-the-news-2-1-Component-one).

  • 01_preprocess.ipynb prepares the article corpus and produces the required embeddings. Note: this notebook downloads the full All the News 2.1 dataset (~8.8 GB) from HuggingFace. If you only want to explore the recommendation step, you can skip this notebook and run 02_recommend.ipynb directly using the pre-computed data/01_processing_output.csv already included in the repository.
  • 02_recommend.ipynb generates recommendation types and provides exploration of top-N recommendations.
  • utils/ — contains keybert_return_embeds.py (a required KeyBERT wrapper that returns (keyword, score, embedding) tuples).
  • data/ — default directory for preprocessed CSV: 01_processing_output.csv; and for HuggingFace dataset download.

• bnrs_analysis/ — code to reproduce the statistical analyses reported in the paper (validation, H1a-c, H2a-c).

  • 01_preprocess.ipynb documents how participant and recommendation-level inputs were prepared for modeling. Note: this notebook requires raw news article data and Qualtrics survey exports that cannot be redistributed due to copyright and privacy restrictions. It is included for transparency but cannot be re-run. The model-ready outputs it produces are provided in data/processed/.
  • 02_models.ipynb fits the mixed-effects ordinal and binary logistic regression models and produces the reported results. This is the notebook to run for replication — it uses the pre-processed data in data/processed/ and is fully self-contained.
  • data/processed/ — model-ready data used by 02_models.ipynb. Contains anonymized participant-level data restricted to the fields required to run the models (model_df.csv, supplement_df.csv, HxRyRz_recs.json, transitions.json).

Setup

This project uses uv for Python dependency management.

Algorithm notebooks (bnrs_algorithm/)

uv sync --extra algorithm
uv run python -m spacy download en_core_web_lg
uv run python -c "import nltk; nltk.download('stopwords')"

Analysis notebooks (bnrs_analysis/)

The analysis notebooks use R via rpy2. You need both the Python dependencies and a working R installation with the required packages.

Python dependencies:

uv sync --extra analysis

R dependencies:

Install R (>= 4.4) and the following packages:

install.packages(c("lme4", "Matrix", "performance", "insight", "ordinal", "MuMIn"))

If you prefer conda for the R setup:

conda install r-base=4.4.3 r-lme4=1.1 r-matrix=1.7 r-performance=0.15 r-insight=1.4 r-ordinal=2023.12 r-mumin=1.48 -c conda-forge

Citation

If you use this code, please cite the paper:

@article{Kiddle2026BisociativePivoting,
  author  = {Kiddle, Rupert and Welbers, Kasper and Kroon, Anne and Trilling, Damian},
  title   = {Bisociative Pivoting as a Method for Moderating User Response to Text Similarity in Next-item News Recommendation},
  journal = {ACM Transactions on Recommender Systems},
  year    = {2026}
}

Contact

Rupert Kiddle — rptkiddle@gmail.com