Revisiting Sentiment and Emoji Signals in Hybrid Recommender Systems: A Case Study

This repository contains a modular Python implementation of a hybrid recommender system enhanced with emoji-based sentiment enrichment and SHAP explanations. It also includes evaluation metrics like AUC, Precision@5, and Recall@5. This work is prepared for submission to ACM RecSys 2025 towards the LBR track.

The work is divided into baseline model and three other hybrid models: emoji_model, tfidf_model and lda_model

Setup Instructions

1. Clone the Repository

git clone https://github.com/antgouri/emoji-recsys.git

cd emoji-recsys

2. Create a Conda Environment and Install Requirements (We prefer to use conda environments only)

conda create -n emosys python==3.10

conda activate emosys

pip install -r requirements.txt

python -m textblob.download_corpora

Run the Full Pipeline with the commands

• python scripts/run_pipeline.py baseline (to run the baseline model)
• python scripts/run_pipeline.py emoji_model (to run the emoji-infused model)
• python scripts/run_pipeline.py tfidf_model (to run the TfIDF model)
• python scripts/run_pipeline.py lda_model (to run the LDA model)

This will:

1. Load and preprocess the data from Musical_Instruments_5.json.
2. Inject sentiment-based emojis to the reviews.
3. Train a LightFM hybrid recommendation model.
4. Generate SHAP explanations and save plots to outputs/.
5. Evaluate the model using AUC, Precision@5, and Recall@5.

Output Artifacts

Artifact	Location
Emoji-injected reviews	csv/emoji_reviews.csv
SHAP summary plot	outputs/shap_summary.png
SHAP raw values	outputs/shap_values.npy
Evaluation results	Printed to terminal

Evaluation Metrics

The recommender is evaluated using:

• AUC (Area Under Curve): How well the model ranks positive vs negative items.
• Precision@5: Proportion of top-5 recommended items that are relevant.
• Recall@5: Proportion of all relevant items recommended in top-5.

PS: For the sake of reproducibility of same results (evaluation metrics) - the LightFM model is making use of the "logistic" loss function instead of the regular warp loss function The randomness of sampling is preserved at every place with random_state=42, random_seed at global level and the number of threads is made non-parallel with a value of 1

Dataset Used

• Source: Amazon Musical Instruments 5-core dataset
• Format: JSON
• Fields Used: reviewerID, asin, overall, reviewText

Dependencies

Listed in requirements.txt:

• lightfm, textblob, pandas, scikit-learn, shap, matplotlib, nltk, cupy-cuda12x (for GPU support), seaborn

Authors

Developed by Dr. Ananth G S and Dr. K. Raghuveer as part for LBR Track for ACM RecSys 2025.

Feedback & Contributions

Feel free to open issues or submit pull requests to improve sentiment detection, support new explainability methods, or add datasets!

Cite our work

@inproceedings{emojiRecSys2025, title={Revisiting Sentiment and Emoji Signals in Hybrid Recommender Systems: A Case Study}, author={Ananth G S, K Raghuveer}, booktitle={Proceedings of the Late-Breaking Results Track of the 19th ACM Conference on Recommender Systems (RecSys '25 LBR)}, year={2025}, publisher={ACM}, url={https://github.com/antgouri/emoji-recsys}, note={Available as a Late-Breaking Result at ACM RecSys 2025} }

Star this repo if you find it useful or insightful!