Research project at the Politecnico di Milano.
Objective: to achieve performance comparable to traditional RecSys, improving user satisfaction through explanations.
If you find this work useful for your research, please cite us using the following .bib entry:
@inproceedings{
author = {Andrea Pisani and
Luca Pagano and
Giuseppe Vitello and
Maurizio {Ferrari Dacrema} and
Paolo Cremonesi},
title = {Blended User Embeddings for Steerable Matrix Factorization},
booktitle = {Proceedings of the 20th {ACM} Conference on Recommender Systems, RecSys
2026, Minneapolis, MN, USA, September 27-October 2, 2026},
publisher = {{ACM}},
year = {2026},
url = {https://doi.org/10.1145/3773078.3841260},
doi = {10.1145/3773078.3841260}
}From HuggingFace
Prototype/
├─ Dataset/ <- Contains processed datasets and embeddings
├─ Decoder/ <- Custom architectures including BlendedALS
├─ LLM/ <- Code for prompts, summaries and embeddings
│ ├─ BLUE_Steer/
├─ optuna/ <- Some of the optuna studies we ran
├─ utils/ <- Various utilities
RecSysFramework/ <- Lab framework with some modifications
Each dataset has the following structure
final/
├─ movie_desc_history/
├─ test_recommendations.npz
├─ train_recommendations.npz
tuning/
├─ movie_desc_history/
├─ test_recommendations.npz
├─ train_recommendations.npz
item_embeddings_X.npz
Where tuning is the train+val dataset, while final is the (train+val) + test dataset.
Specifically, train_recommendations is the training dataset, while test_recommendations is for evaluation.
Inside the dataset folder there will be item_embeddings where X will be the embedding model.
Inside movie_desc_history there are the data generated from the user taste summaries.
Inside there will be a user_prompts_DESC.parquet file which is the parquet containing the prompt passed to the LLM that generates the summaries.
The subfolders are the models used to generate the summaries.
Inside there will be a file like user_desc_from_LLM_... containing the response, i.e., the summaries and then the embeddings of the various models
Only the ML Small dataset is detailed here
- create_user_description.py: creates the prompt where given the user history and associated metadata like textual descriptions, we ask to build a summary. We get a .parquet file.
- user_summary_llama.cpp.py: interfaces with the llama.cpp backend to generate summaries. There is an option to enable or disable the enhancing prompt which is used only for local summaries to improve response style (detailed in the accompanying publication).
- gemini_api_cache.py: same as above but using the gemini API
Common to all datasets
- embedder_ollama.py: to create embeddings