Compressed ELSA: Efficient Sparse Representations for Large-Scale Recommendation

Official repository for the paper "Efficient Sparse Representations for Large-Scale Recommendation", accepted to the THE ACM WEB CONFERENCE 2026 (WWW 2026) - Short Paper Track.

Overview

Compressed ELSA is a sparse, scalable variant of the ELSA collaborative-filtering autoencoder. It learns high-dimensional sparse item embeddings directly during training, significantly reducing model size and speeding up retrieval while preserving recommendation quality.

This enables:

• 10×–100× smaller smaller embeddings with minimal accuracy loss, competitive with strong baselines on several datasets
• Faster inference via sparse matrix–vector operations
• Interpretable latent spaces, where dimensions naturally correspond to item segments

Live Demo

Our interactive demo showcases additional experiments, ablations, and visualizations that did not fit into the manuscript.

You can also explore:

• Segment-level recommendation
• Interpretability of the learned sparse latent factors

The source code for the demo lives in the demo branch.

Repository Layout

run.py                Main entry point (training / evaluation controller)
recommenders/
    baselines.py        Baseline recommender models for comparison
    elsa_models.py      ELSA variants (dense, hybrid, compressed-sparse)
experiments/          Experiment runner scripts and configs
results/              Metrics, logs, checkpoints, run artifacts
_datasets/            Download / load / parse / preprocess utilities + data

Environment Setup

python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
# or manual:
pip install numpy scipy torch keras scikit-learn recpack tqdm pandas implicit

Example Usage (Training)

From repository root, run, e.g.,

python experiments/experiment_compressed_elsa.py --dataset ml20m --factors 4096 --batch_size 1024 --max_output 20000 --decay_strategy Exponential --vals "0 1024 512 256 128 64 32 16" --lth True

Check individual experiments for available arguments.

Citation

(To be added once the camera-ready version is finalized.)

License

This repository is licensed under CC BY-NC 4.0. See the LICENSE file for details.