Explaining Neural Matrix Factorization with Gradient Rollback

Authors: Carolin Lawrence, Timo Sztyler, Mathias Niepert4987-4995

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We also conduct experiments which show that gradient rollback provides faithful explanations for knowledge base completion and recommender datasets. An implementation1 and an appendix2 are available.
Researcher Affiliation Industry Carolin Lawrence, Timo Sztyler, Mathias Niepert NEC Laboratories Europe, Kurf ursten-Anlage 36, 69115 Heidelberg, Germany {carolin.lawrence, timo.sztyler, mathias.niepert}@neclab.eu
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes An implementation1 and an appendix2 are available. 1https://github.com/carolinlawrence/gradient-rollback
Open Datasets Yes We report results on three datasets: two knowledge base completion (NATIONS (Kok and Domingos 2007), FB15K-237 (Toutanova et al. 2015)) and one recommendation dataset (MOVIELENS (Harper and Konstan 2015)).
Dataset Splits No The paper does not provide specific percentages or counts for training/validation splits. It mentions: "Since retraining is costly, we only explain the top-1 prediction for a set of 100 random test triples for both FB15K-237 and MOVIELENS. For NATIONS we use the entire test set." and "We fix all random seeds and use the same set of negative samples during (re-)training to avoid additional randomization effects."
Hardware Specification Yes For example, FB15K-237 contains 270k training triples and training one model with TF2 on a RTX 2080 Ti GPU takes about 15 minutes.
Software Dependencies No The paper mentions "TF2" implying TensorFlow 2.x, but does not provide a specific version number. It does not list other software dependencies with version numbers.
Experiment Setup No The paper states: "Statistics and hyperparameter settings are in the appendix." and "We fix all random seeds and use the same set of negative samples during (re-)training to avoid additional randomization effects." While the latter is a setup detail in the main text, specific hyperparameter values are deferred to the appendix and not present in the main text.