A Binomial / Gaussian Diffusion Process for 1D Data for Recommendation
A Binomial / Gaussian Diffusion Process for 1D Data for Recommendation
- Developed by: Anonymized
- Model type: Recommendation model
- License: unknown
(i) non-sequential (ii) Top-K recommendation with (iii) binary feedback
Recommendation setups that violate assumptions (i)-(iii).
Significant research has explored bias and fairness issues with language models (see, e.g., Li et al. (2021)). Predictions generated by the model may for example unfairly favor more active users and more popular items.
MovieLens1M, MovieLens10M, MovieLens25M, Netflix
(i) filter out users with fewer than five items, and items with fewer than five interactions (ii) Users are split into train, validation and test sets (0.8 / 0.1 / 0.1), with the training employing the entire history. For validation and test sets, a partial history is fed to the recommender, with a held-out set being used to evaluate the resulting recommendation. [a.k.a. strong generalization] (iii) binarize feedback: ratings above 3 are encoded as 1, and zero otherwise.
Recall@20, Recall@50, NDCG@100
Dataset | Model | Recall@20 | Recall@50 | NDCG@100 |
---|---|---|---|---|
MovieLens25M | Random | 0.13 | 0.30 | 0.24 |
Popularity | 16.63 | 24.43 | 19.69 | |
\textbf{RecFusion} | 33.21 | 45.44 | 37.31 | |
CODIGEM | 34.05 | 45.84 | 37.90 | |
MultVAE | 35.12 | 48.09 | 39.12 | |
EASE | 40.02 | 52.71 | 43.84 | |
Netflix | Random | 0.18 | 0.32 | 0.31 |
Popularity | 11.73 | 17.48 | 15.89 | |
CODIGEM | 25.54 | 33.48 | 29.08 | |
\textbf{RecFusion} | 29.68 | 37.63 | 32.87 | |
MultVAE | 31.61 | 40.61 | 35.23 | |
EASE | 36.19 | 44.49 | 39.35 |
- Hardware Type: More information needed
- Hours used: ~ 200 (including hyperparameter tuning)
- Cloud Provider: Azure
- Compute Region: West Europe
- Carbon Emitted: ~ 35 kg CO2 eq.
Carbon emissions were estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).
Either one of MLP, MLP+T, MLP+Var, U-Net1D, U-Net2D
Databricks ML clusters version 13 ML with sparktrials
BibTeX:
Anonymized
APA:
Anonymized
Anonymized