Repository for Multi-Resolution Diffusion for Privacy Sensitive Recommender Systems

Best hyperparameter results in paper

Dataset	ML-100k			ALB			ML-1M			ADM
Recommender Model	SVD	MLP	NeuMF	SVD	MLP	NeuMF	SVD	MLP	NeuMF	SVD	MLP	NeuMF
Best trial number	223	193	44	69	91	67	76	20	4	38	40	22
Recall@10 score	0.3924	0.3839	0.232	0.339	0.3246	0.3225	0.3722	0.3595	0.1026	0.0651	0.0868	0.0234
SDRM batch size	550	810	190	370	530	820	720	160	830	930	270	850
SDRM lr	2.10E-05	5.20E-05	2.80E-05	3.20E-05	3.90E-05	5.90E-05	5.90E-05	9.80E-05	5.00E-06	1.00E-06	6.30E-05	1.30E-05
SDRM epochs	265	200	15	5	200	485	395	15	140	60	45	185
MLP hidden layers*	2	0	4	2	0	2	2	1	1	1	1	5
SDRM timesteps	83	58	138	68	43	33	23	78	178	163	38	93
VAE batch size	780	50	870	420	340	720	190	270	540	380	310	290
VAE hidden layer neurons	930	40	1000	70	550	450	600	490	430	210	20	40
MLP latent neurons	830	40	950	20	400	400	150	340	330	160	20	40
VAE lr	0.0006	0.0034	0.001	0.0042	0.001	0.004	0.0066	0.0002	0.0009	0.0011	0.0035	0.0014
Noise variance diminisher	1	1	0.2	0.5	0.2	0.3	0.5	1	1	0.3	0.7	1

* Number of hidden layers excludes the input and output layers

Usage examples

To recreate the results in the paper, put the best result hyperparamters mentioned in the table in a script similar to below.

MovieLens 100k

Augmenting using SVD

python main.py --dataset ml-100k --model svd --augment-training-data --SDRM-epochs 265 --SDRM-batch-size 550 --SDRM-lr 0.000021 --SDRM-timesteps 83 --SDRM-noise-variance-diminisher 1 --MLP-hidden-layers 2 --VAE-batch-size 780 --VAE-hidden-layer-neurons 930 --MLP-latent-neurons 830 --VAE-lr 0.0006

Amazon Digital Music

No Augmentation using MLP

python main.py --dataset adm --model mlp --augment-training-data --SDRM-batch-size 270 --SDRM-lr 0.000063 --SDRM-epochs 45 --MLP-hidden-layers 1 --SDRM-timesteps 38 --SDRM-noise-variance-diminisher 0.7 --VAE-batch-size 310 --VAE-hidden-layer-neurons 20 --MLP-latent-neurons 20 --VAE-lr 0.0035

---

Results can also be reproduced by specifying the arguments with the result_dict in main.py on line 113

Python version 3.8 or greater was used.

hyperparameter_search_results.7z in the data folder contains all the raw hyperparameter search results and optuna study objects.

Note: This repository contains all the code used to create the results in Multi-Resolution Diffusion for Privacy Sensitive Recommender Systems