Implementation and reproducible experiments behind the research paper "Improving Sequential Recommendation with LLMs", which substantially extends our RecSys paper. The paper is uploaded to Arxiv. Please cite as follows:
@misc{boz2024improving,
title={Improving Sequential Recommendations with LLMs},
author={Artun Boz and Wouter Zorgdrager and Zoe Kotti and Jesse Harte and Panos Louridas and Dietmar Jannach and Marios Fragkoulis},
year={2024},
eprint={2402.01339},
archivePrefix={arXiv},
primaryClass={cs.IR}
}
Implementation and reproducible experiments behind the research paper "Leveraging Large Language Models for Sequential Recommendation" published in RecSys'23 Late-Breaking Results track.
To navigate to the implementation, go to the RecSys23 paper release of this repository.
Please cite as follows.
@inproceedings{Harte2023leveraging,
author = {Harte, Jesse and Zorgdrager, Wouter and Louridas, Panos and Katsifodimos, Asterios and Jannach, Dietmar and Fragkoulis, Marios},
title = {Leveraging Large Language Models for Sequential Recommendation},
year = {2023},
isbn = {979-8-4007-0241-9/23/09},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3604915.3610639},
doi = {10.1145/3604915.3610639},
booktitle = {Proceedings of the 17th ACM Conference on Recommender Systems},
numpages = {7},
location = {Singapore, Singapore},
series = {RecSys '23}
}
In order to develop code in this repository, you'll need poetry. In case you don't have poetry installed, follow the steps in the steps in the official documentation.
This will activate a shell within the current poetry environment. If the environment does not exist, it will be automatically created.
poetry shellAlternatively, you can prefix your commands with
poetry run
which will have the same effect as executing commands within the poetry environment.
This command will also install the llm-sequential-recommendations package in editable mode. This means that any changes made in the code will be automatically reflected, as long as they reside in directories that existed during the installation. If at any point you get errors that do not make sense, feel free to run the command again.
poetry installWe organized the repository by distinguishing between dataset-related code (e.g. beauty), implementation code (inside main), notebooks for analysis and visualization (inside notebooks), and results (in results). We explain our dataset-related code in a separate README.md in the corresponding directory.
For now we have included all code we used to process Beauty in the beauty directory with a separate README.md. Regarding the Delivery Hero dataset, we are discussing with our organization to make it available.
The implementation code in main first of all contains data and eval. The former contains the implementation of our SessionDataset, which is a convenient object to group together all the code and data related to a single dataset, including the train and test set, but also the train and test set of the validation folds. The latter contains the implementation of each of the metrics and evaluation.py, which converts the recommendations and ground-truths into the format expected by the metrics, evaluates the recommendation using the metrics, and subsequently provides a view of the evaluation on all metrics.
Furthermore, main contains the implementations of each recommendation model in their respective directory inside the main directory. The exception here are LLMSeqSim, LLMSeqPrompt, and the hybrids. LLMSeqSim and LLMSeqPrompt are grouped together under llm_based, since both of these models require the utilities in embedding_utils. Both models require the product_embeddings_openai.csv.gzip file created by create_embeddings.ipynb. This notebook in turn uses openai_utils.py, which requires the OpenAI API key to be set in key.txt. We added the embeddings created by create_embeddings.ipynb for the Beauty dataset in the beauty directory. Note that LLMSeqSim is implemented in main/llm_based/similarity_model, and LLMSeqPrompt is implemented in main/llm_based/prompt_model.
In addition, the hybrids (EmbeddingEnsemble, LLMSeqSim&Sequential, LLMSeqSim&Popularity) are grouped together under main/hybrids because they all require the utilities and properties placed in utils.py and properties.py, respectively. EmbeddingEnsemble, LLMSeqSim&Sequential, LLMSeqSim&Popularity are implemented in embedding-ensemble.py, popularity-based-hybrid.py, and property-injected-embedding.py, correspondingly. To run the hybrids you can use the command python <hybrid-filename>. You need to have an MLflow username and password set in the respective environment variables MLFLOW_TRACKING_USERNAME and MLFLOW_TRACKING_PASSWORD so that the individual model configurations can be retrieved.
All models except LLMSeqSim and LLMSeqPrompt implement the interface of abstract_model. The train method accepts a pd.DataFrame containing SessionId and ItemId as columns. The predict method accepts a dictionary mapping SessionId to a 1-dimensional np.ndarray of items, sorted according to time. The predict method returns an other dictionary mapping SessionId to an other 1-dimensional np.ndarray of recommended items, sorted descendingly on confidence (so the most confidence is on position 0). For more implementation details on the models, refer to online_material.md.
Running experiments using the models and a given dataset is done in run_experiments.ipynb. Make sure to use the correct poetry environment (see #Installation) when running the notebook. You can set DATASET_FILENAME to the path of the dataset pickle. The product embeddings of the dataset are stored with Git LFS and should be downloaded with git lfs pull -- <embeddings_filename>. EXPERIMENTS_FOLDER is the directory to which the recommendations and configurations are persisted, and from which the evaluation retrieves all recommendation pickles to evaluate. We hardcoded the optimal configurations returned by our hypersearch. In addition, we persisted the weights of the neural models with the top-performing configurations to ensure reproducilibity.