Dargk Team Submission - RecSys 2025 Challenge

Welcome to the official repository of the Dargk team submission for the RecSys 2025 Challenge. This repository contains the full pipeline and implementation of our model, BEHAV-E: Behavioral Embedding via Hybrid Action Variational Encoder.

Our approach is centered on modeling user behavior through a hybrid representation combining semantic, categorical, and temporal information, and training via a contrastive learning framework.

---

🧠 Model Overview

BEHAV-E: Behavioral Embedding via Hybrid Action Variational Encoder

BEHAV-E is a representation learning model designed to encode complex and multi-faceted user behavior. The model processes various user actions such as:

• Product buys
• Product add-to-cart
• Product remove-from-cart
• URL visits
• Search queries

Key modeling elements include:

• Kernel Density Estimation (KDE) to capture temporal distribution of actions.
• LSTM-based Autoencoder to embed user search queries.
• Shared Embedding Bags with linear transformations to efficiently represent items, categories, and URLs.
• Variational Encoder to output user behavior embeddings using reparameterization trick.
• Contrastive Learning (InfoNCE loss) with dual models to learn discriminative embeddings.

During inference, we concatenate the mean embeddings from two BEHAV-E models as an ensemble strategy.

---

📁 Repository Structure

• EmbeddingAnalysis: contains the code for reproduce the analysis as well as more details than the ones presented on the paper.
• data/: Folder where the decompressed dataset should be placed
• 01-TextEncodeTrain.py: Trains the LSTM autoencoder for text embeddings
• 02-TextEncodeProcess.py: Encodes the search text using the trained autoencoder
• 03-Polars_DS_enc_search.py: Preprocesses data: one row per client format
• 04-Emb-8ShortVAECATURLFastNAEMA.py: Trains BEHAV-E using EMA (Exponential Moving Average)
• 04-Emb-8ShortVAECATURLFastNAOneCycleLarge.py: Trains BEHAV-E using One Cycle LR schedule
• 05-Emb-8ShortVAECATURLFastNAEMA_gen.py: Generates embeddings using EMA-trained model
• 05-Emb-8ShortVAECATURLFastNAOneCycleLarge_gen.py: Generates embeddings using OneCycle-trained model
• 06-Merger.py: Concatenates embeddings for final submission
• environment.yml: Conda environment definition for Windows
• Dockerfile: Dockerfile to build a docker container for out model.
• README.md: This file

Execution Note: Files should be executed sequentially from 01 to 06. Files with the same prefix number (e.g., both 04-*) can be run in any order.

---

⚙️ Setup

There are two ways to setup teh environment. The first one is in the local environment through conda with the provided environment.yml file. This has been tested only on a Windows environment. The second way is using a docker container with the provided Dockerfile.

Notice that this project only depends on the data, no pretrained model is required to run the pipeline end-to-end. However, we provide our trained models to reproduce the embeddings submitted to the challenge. The link is provided below.

1. Clone the repository:

git clone https://github.com/<your-org>/recsys2025-dargk.git
cd recsys2025-dargk

💻 Local

1. Set up the conda environment:

conda env create -f environment.yml
conda activate behav-e

🐳 Docker (Recommended)

Follow these instructions to build and run the RecSys2025 Docker container with GPU support.

1. Build the Docker Image

docker build -t recsys2025 .

2. Create the Docker Container
   1. Windows

   docker container create -i -t -v "%CD%":/recsys2025 --gpus=all --name recsys2025 recsys2025

   2. Linux/macOS

   docker container create -i -t -v "$PWD":/recsys2025 --gpus=all --name recsys2025 recsys2025

3. Start and Attach to the Container

docker container start --attach -i recsys2025

💾 Preparing the dataData

To run the system is necesary to download the dataset, which is compressed in a file calles ubc_data.tar.gz. This file must be extracted into the data/ directory.

A detailed description of the dataset can be found at the challenge Web site. It containes six parquet files:

• product_buy.parquet

  • client_id (int64): Numeric ID of the client (user).
  • timestamp (object): Date and time of the event in the format YYYY-MM-DD HH:mm:ss.
  • sku (int64): Numeric ID of the item.

• add_to_cart.parquet

  • client_id (int64): Numeric ID of the client (user).
  • timestamp (object): Date and time of the event in the format YYYY-MM-DD HH:mm:ss.
  • sku (int64): Numeric ID of the item.

• remove_from_cart.parquet

  • client_id (int64): Numeric ID of the client (user).
  • timestamp (object): Date and time of the event in the format YYYY-MM-DD HH:mm:ss.
  • sku (int64): Numeric ID of the item.

• product_properties.parquet

  • sku (int64): Numeric ID of the item.
  • category (int64): Numeric ID of the item category.
  • price (int64): Numeric ID of the item's price bucket.
  • embedding (object): A textual embedding of a product name, compressed using the product quantization method.

• page_visit.parquet

  • client_id (int64): Numeric ID of the client.
  • timestamp (object): Date and time of the event in the format YYYY-MM-DD HH:mm
  • url (int64): Numeric ID of a visited URL. The explicit information about what (e.g., which item) is presented on a particular page is not provided.

• search_query.parquet

  • client_id (int64): Numeric ID of the client.
  • timestamp (object): Date and time of the event in the format YYYY-MM-DD HH:mm:ss.
  • query (object): The textual embedding of the search query, compressed using the product quantization method. ‍ The dataset also containes two more directories related to the task.

• input directory: This directory stores a NumPy file containing a subset of 1,000,000 client_ids for which Universal Behavioral Profiles should be generated:

  • relevant_clients.npy

• target directory: This directory stores the labels for propensity tasks. For each propensity task, target category names are stored in NumPy files:

  • propensity_category.npy: Contains a subset of 100 categories for which the model is asked to provide predictions
  • popularity_propensity_category.npy: Contains popularity scores for categories from the propensity_category.npy file. Scores are used to compute the Novelty measure.
  • propensity_sku.npy: Contains a subset of 100 products for which the model is asked to provide predictions
  • popularity_propensity_sku.npy: Contains popularity scores for products from the propensity_sku.npy file. These scores are used to compute the Novelty measure.
  • active_clients.npy: Contains a subset of relevant clients with at least one product_buy event in history (data available for the participants). Active clients are used to compute churn target.

---

🚀 Pipeline Execution

Follow the steps below to reproduce the full embedding generation and submission process.

1. Train the LSTM Autoencoder for Search Query Embeddings

python 01-TextEncodeTrain.py

2. Encode the Search Texts

python 02-TextEncodeProcess.py

3. Preprocess the Dataset to Generate One Row per User

python 03-Polars_DS_enc_search.py

4. Train the BEHAV-E Models

• Exponential Moving Average (EMA) Variant:

python 04-Emb-8ShortVAECATURLFastNAEMA.py

• One Cycle Learning Rate Variant:

python 04-Emb-8ShortVAECATURLFastNAOneCycleLarge.py

5. Generate User Embeddings from Trained Models

• From EMA-trained model:

python 05-Emb-8ShortVAECATURLFastNAEMA_gen.py

• *From OneCycle-trained model:

python 05-Emb-8ShortVAECATURLFastNAOneCycleLarge_gen.py

6. Merge the Generated Embeddings for Final Submission

python 06-Merger.py

⚠️ Note: All scripts with the same prefix number (e.g., 04-, 05-) can be run in any order.

🧪 Inference Details

At inference time:

• BEHAV-E uses the mean of the latent distribution, i.e., $$\mu = \mathbb{E}[z]$$, instead of sampling from $$\mathcal{N}(\mu, \sigma^2)$$.
• Dropout is disabled to ensure deterministic outputs.
• The final user embedding is the concatenation of the embeddings produced by both trained BEHAV-E models (EMA and OneCycle variants). Although both models are trained on the same data, this concatenation acts as a lightweight ensemble and improves robustness.

---

Submission

The pretrained models used for generate our submission can be downloaded from an external repository.

---

🙌 Acknowledgements

We would like to thank the organizers of the RecSys 2025 Challenge for providing a valuable dataset and a well-designed competition platform.

---

📝Cite us!

@inproceedings{10.1145/3758126.3758130,
   author = {Rodriguez, Juan Manuel and Tommasel, Antonela},
   title = {BEHAV-E! You are Not Just a Number to Us, but an {$\mathbb{R}^{2048}$} Embedding},
   year = {2025},
   isbn = {9798400720994},
   publisher = {Association for Computing Machinery},
   address = {New York, NY, USA},
   url = {https://doi.org/10.1145/3758126.3758130},
   doi = {10.1145/3758126.3758130},
   booktitle = {Proceedings of the Recommender Systems Challenge 2025},
   pages = {16–20},
   numpages = {5},
   series = {RecSysChallenge '25}
}

📫 Contacts

For inquiries, collaboration, or questions regarding this submission, please contact the Dargk Team at:

📧 Antonela Tommasel (antonela.tommasel@isistan.unicen.edu.ar)

📧 Juan Manuel Rodriguez (jmro@cs.aau.dk)