Ordinal Non-negative Matrix Factorization for Recommendation
Authors: Olivier Gouvert, Thomas Oberlin, Cédric Févotte
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We report recommendation experiments on explicit and implicit datasets, and show that Ord NMF outperforms Be Po F and PF applied to binarized data. |
| Researcher Affiliation | Academia | 1IRIT, Université de Toulouse, CNRS, France 2ISAESUPAERO, Université de Toulouse, France. |
| Pseudocode | Yes | Algorithm 1 CAVI for IG-Ord NMF. |
| Open Source Code | Yes | All the Python codes are available on https://github.com/Oligou/OrdNMF. |
| Open Datasets | Yes | Movie Lens (Harper & Konstan, 2015). This dataset contains the ratings of users on movies on a scale from 1 to 10. These explicit feedbacks correspond to ordinal data. We consider that the class 0 corresponds to the absence of a rating for a couple user-movie. The histogram of the ordinal data is represented in blue on Figure 4. We pre-process a subset of the data as in (Liang et al., 2016), keeping only users and movies that have more than 20 interactions. We obtain U = 20k users and I = 12k movies. Taste Profile (Bertin-Mahieux et al., 2011). This dataset, provided by the Echo Nest, contains the play counts of users on a catalog of songs. As mentioned in the introduction, we choose to quantize these counts on a predefined scale in order to obtain ordinal data. |
| Dataset Splits | No | The paper specifies a train/test split ("the train set contains 80%... the test set contains the remaining 20%") but does not explicitly mention a separate validation set. |
| Hardware Specification | Yes | The computer used for these experiments was a Mac Book Pro with an Intel Core i5 processor (2,9 GHz) and 16 Go RAM. |
| Software Dependencies | No | The paper mentions that the code is in Python, but does not specify version numbers for Python itself or any specific libraries (e.g., PyTorch, TensorFlow, scikit-learn) that would be needed for replication. |
| Experiment Setup | Yes | For all models, we select the shape hyperparameters αW = αH = 0.3 among {0.1, 0.3, 1} (Gopalan et al., 2015). The number of latent factors is chosen among K {25, 50, 100, 150, 200, 350} for the best NDCG score with threshold s = 8 for the Movie Lens dataset, and s = 1 for the Taste Profile dataset. All the algorithms are run 5 times with random initializations and are stopped when the relative increment of the expected lower bound (ELBO) falls under τ = 10-5. |