HoloMambaRec

HoloMambaRec is a minimal, single-file reference implementation of holographic binding with Mamba state space models for sequential recommendation. It includes data loaders for Amazon Beauty and MovieLens-1M, model definitions (HoloMamba, SASRec, GRU4Rec, and an item-only ablation), and scripts for benchmarking, ablations, compression profiling, and lightweight grid search.

Repository contents

• HoloMambaRec.py — end-to-end script (data prep, models, training loops, evals, and plotting).

Setup

Python 3.9+ and a CUDA GPU are recommended.

python3 -m venv .venv
source .venv/bin/activate
pip install torch mamba-ssm causal-conv1d pandas numpy matplotlib tqdm requests
# Optional (if you want parity with the Colab snippet in the code header):
# pip install datasets transformers google-generativeai

Running the script

The main entry point is HoloMambaRec.py. It will download datasets on first run into data/amazon-beauty and data/ml-1m.

python HoloMambaRec.py

What it does:

• Runs the benchmark loop on Amazon-Beauty and ML-1M with HoloMamba, SASRec, and GRU4Rec.
• Runs the binding vs. item-only ablation.
• Runs a grid search and writes grid_search_results.csv (set GRID_MAX_TRIALS=0 or a small integer to limit/skip it).
• Runs compression/latency evaluations and saves plots.

Notes:

• The full run is compute-heavy; use a GPU and consider trimming epochs or commenting out sections in __main__ for quicker smoke tests.
• Figures and CSVs are written to the repository root (e.g., learning_curve_amazon-beauty_hr.png, compression_runtime_ml-1m.png).

Customization tips

• Adjust global defaults in GLOBAL_CONFIG (e.g., d_model, n_layers, batch_size, epochs, use_compression).
• Modify dataset-specific overrides inside run_benchmark and run_compression_evals.
• To evaluate only cold-start behavior or compression, call evaluate_cold_start or run_compression_evals directly from a notebook after constructing your loaders.

Attribution

If you use this codebase, please reference the HoloMambaRec project and include links back to this repository.